/h/ - Hentai » Thread #7379025

[1157 / 309 / ?]

/hdg/ - Hentai Diffusion General #323

Anonymous Tue 16 May 2023 20:37:40 No.7379025 View View Reply Original Report

Quoted By: >>7379038 >>7379042 >>7379058 >>7379066 >>7379182 >>7379449 >>7380136 >>7382186

Capybara Edition

Old: >>7376137

>RESOURCES
Wiki: https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki
LoRAs: https://gitgud.io/gayshit/makesomefuckingporn | https://github.com/derrian-distro/LoRA_Easy_Training_Scripts
Tags: https://danbooru.donmai.us/wiki_pages/tag_groups
ControlNet: https://rentry.org/dummycontrolnet
Animation: https://rentry.org/AnimAnon
Wildcards: https://rentry.org/NAIwildcards
4chanX Catbox userscript: https://gist.github.com/catboxanon/ca46eb79ce55e3216aecab49d5c7a3fb
OP Template/Logos: https://pastebin.com/0atr0fEc | https://files.catbox.moe/xk99gy.png | https://files.catbox.moe/35bdt5.gif

Anonymous

Anonymous Tue 16 May 2023 20:39:26 No.7379036 Report

Quoted By: >>7379043 >>7379048 >>7379053 >>7379055 >>7379057 >>7379066 >>7379122 >>7379200

Since it's been brought up in the previous threads, are there any resources you would like to see added to the OP?
What do you feel is missing from it?

Anonymous

Anonymous Tue 16 May 2023 20:40:07 No.7379038 Report

Quoted By: >>7379066

>>7379025
it just started...

Anonymous

Anonymous Tue 16 May 2023 20:41:18 No.7379042 Report

Quoted By: >>7379044 >>7379066

>>7379025
>capybara not flaming
one job op

Anonymous

Anonymous Tue 16 May 2023 20:41:22 No.7379043 Report

Quoted By: >>7379066

>>7379036
wednesday addams guide

Anonymous

Anonymous Tue 16 May 2023 20:41:58 No.7379044 Report

Quoted By:

>>7379042
Why would you want him on fire?

Anonymous

Anonymous Tue 16 May 2023 20:43:03 No.7379048 Report

Quoted By: >>7379053 >>7379055 >>7379066

>>7379036
Install guide for brainlets with models/vae shit explained
Something like voldy's guide but not severely outdated

Anonymous

Anonymous Tue 16 May 2023 20:43:54 No.7379049 Report

Quoted By:

bwos how do I avoid a distinct color change when using multidiff? it seems to desaturate and brighten the image quite a bit

Anonymous

Anonymous Tue 16 May 2023 20:44:48 No.7379053 Report

Quoted By:

>>7379036
>>7379048
Yeah this.
That one obtuse guide explaining lora parameters would be okay too if it was cleaned and transcribed onto a normal platform (not rentry).
I dare not suggest a lora guide because that'd be a disaster.

Anonymous

Anonymous Tue 16 May 2023 20:45:41 No.7379055 Report

Quoted By: >>7379062 >>7379065 >>7379069 >>7380230

>>7379036
>>7379048
if 90% of the old guides suck/are outdated why don't we just make new guides that aren't retarded?

Anonymous

View Same Google ImgOps iqdb SauceNAO butt.png, 2MiB, 1024x1536

Anonymous Tue 16 May 2023 20:45:48 No.7379056 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO 9996490989765701-2962968559-kill (...).png, 2MiB, 1600x2200

Anonymous Tue 16 May 2023 20:45:57 No.7379057 Report

Quoted By: >>7379075

>>7379036
maybe other diffusion boards would be nice

Anonymous

View Same Google ImgOps iqdb SauceNAO 00526-2279576463, DPM++ 2M Karra (...).jpg, 608KiB, 768x1280

Anonymous Tue 16 May 2023 20:46:50 No.7379058 Report

Quoted By: >>7379072

>>7379025
naisu

Anonymous

View Same Google ImgOps iqdb SauceNAO 00474-3379097206.jpg, 800KiB, 1993x4100

Anonymous Tue 16 May 2023 20:47:13 No.7379060 Report

Quoted By:

Honestly DPM++ 2M SDE/Karras seems to be pretty impressive, but you have to discard the last step to work fine.

Same speed as DPM++ 2M/Karras.

https://files.catbox.moe/ilfxhq.png

Anonymous

Anonymous Tue 16 May 2023 20:47:21 No.7379062 Report

Quoted By:

>>7379055
Because it'd take time and effort to write

Anonymous

View Same Google ImgOps iqdb SauceNAO 08028-3413667272-1girl, black ha (...).png, 2MiB, 1024x1024

Anonymous Tue 16 May 2023 20:47:47 No.7379063 Report

Quoted By:

nearly burst into treats when the larger one materialized

Anonymous

Anonymous Tue 16 May 2023 20:48:28 No.7379065 Report

Quoted By:

>>7379055
Perfect timing, the LORA guides at the top of the thread, are they the most contemporary instructions for some one with average SD skills, or do folks have a recommendation?

Anonymous

Anonymous Tue 16 May 2023 20:48:44 No.7379066 Report

Quoted By: >>7379074 >>7379085

>>7379025
>>7379036
>>7379038
>>7379042
>>7379043
>>7379048

I haven't been into SD since January, what's the meta for models? Are people still using NAI? Did WD 1.4 ever come out?

Anonymous

View Same Google ImgOps iqdb SauceNAO merge pen on face text jpg.jpg, 216KiB, 1024x1536

Anonymous Tue 16 May 2023 20:49:21 No.7379067 Report

Quoted By: >>7379831

meanies

Anonymous

Anonymous Tue 16 May 2023 20:50:28 No.7379069 Report

Quoted By:

>>7379055
I feel like this community is pretty fragmented compared to other ones. We can barely get people to bake most days.
A few autists made the early ones and then nobody else cared (neither did they tbf) to do it. That's why most of this stuff is pre-February.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00022-3432119541.png, 2MiB, 1024x1536

Anonymous Tue 16 May 2023 20:51:31 No.7379072 Report

Quoted By: >>7379081

>>7379058
great gen, amazing reflection. catbox?

Anonymous

View Same Google ImgOps iqdb SauceNAO 1680891230827835.png, 2MiB, 1080x1280

Anonymous Tue 16 May 2023 20:53:08 No.7379073 Report

Quoted By: >>7379076

This bitch really taking off her panties in a 7/11!!

Anonymous

Anonymous Tue 16 May 2023 20:53:19 No.7379074 Report

Quoted By: >>7379130 >>7382066

>>7379066
People only use NAI as a base model to train loras on, never for actual prompting
WD 1.5 came out a few days ago and although I haven't tried it myself other posters didn't seem to impressed by it

Nowadays people use lots of various custom mixes downloaded from either Civitai or Huggingface, you could browse the sites and see if there's something you think looks interesting. Personally I mostly use BasedMix64, Mokoumix and Butter. All of them are on Huggingface

Anonymous

View Same Google ImgOps iqdb SauceNAO 9996490989765674-801391981-killj (...).png, 2MiB, 1600x2600

Anonymous Tue 16 May 2023 20:53:31 No.7379075 Report

Quoted By: >>7379843

>>7379057

Anonymous

Anonymous Tue 16 May 2023 20:54:07 No.7379076 Report

Quoted By:

>>7379073
>those hands
she works at a spooky mansion doesn't she

Anonymous

View Same Google ImgOps iqdb SauceNAO 00218-3181864063, DPM++ 2M Karra (...).jpg, 593KiB, 768x1280

Anonymous Tue 16 May 2023 20:57:19 No.7379081 Report

Quoted By:

>>7379072
https://files.catbox.moe/hen3ln.jpg

Anonymous

View Same Google ImgOps iqdb SauceNAO marble.png, 1MiB, 968x968

Anonymous Tue 16 May 2023 21:00:26 No.7379085 Report

Quoted By: >>7379130

>>7379066
I think current meta revolves around interesting LoRAs.

Anonymous

Anonymous Tue 16 May 2023 21:01:54 No.7379086 Report

Quoted By: >>7379087

i haven't updated the ui for months, is it safe yet or still fucked

Anonymous

Anonymous Tue 16 May 2023 21:02:46 No.7379087 Report

Quoted By:

>>7379086
nah don't do it

Anonymous

Anonymous Tue 16 May 2023 21:10:27 No.7379096 Report

Quoted By: >>7379104 >>7379118

>>7378916
Where's the sinkai lora?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00129-887102503.png, 1MiB, 864x1152

Anonymous Tue 16 May 2023 21:16:20 No.7379101 Report

Quoted By:

Anonymous

Anonymous Tue 16 May 2023 21:18:24 No.7379104 Report

Quoted By: >>7379108 >>7379109

>>7379096
here https://files.catbox.moe/oofnvp.safetensors
I jus baked it today

Anonymous

Anonymous Tue 16 May 2023 21:24:53 No.7379108 Report

Quoted By:

>>7379104
Many thanks based anon

Anonymous

View Same Google ImgOps iqdb SauceNAO 61e42dee2449c00679c862d61131165a (...).png, 2MiB, 1024x1536

Anonymous Tue 16 May 2023 21:25:39 No.7379109 Report

Quoted By:

>>7379104
nta but thank u

Anonymous

Anonymous Tue 16 May 2023 21:33:51 No.7379117 Report

Quoted By: >>7379119 >>7379121 >>7379124

Anons please give me an idea for a prompt and i'll run a batch of 200 while taking my bath

Anonymous

Anonymous Tue 16 May 2023 21:34:57 No.7379118 Report

Quoted By:

>>7378916
bases stinky bitch enjoyer
>>7379096
more like stinkai

Anonymous

Anonymous Tue 16 May 2023 21:35:39 No.7379119 Report

Quoted By:

>>7379117
obviously prompt something bathing related

Anonymous

Anonymous Tue 16 May 2023 21:36:08 No.7379121 Report

Quoted By:

>>7379117
the james joyce letters

Anonymous

Anonymous Tue 16 May 2023 21:36:45 No.7379122 Report

Quoted By:

>>7379036
at least link to the other threads

Anonymous

View Same Google ImgOps iqdb SauceNAO catbox_glcapz.png, 887KiB, 768x1152

Anonymous Tue 16 May 2023 21:36:51 No.7379123 Report

Quoted By: >>7379124 >>7379137

Here you go. I'm working on this prompt right now.

Anonymous

Anonymous Tue 16 May 2023 21:37:56 No.7379124 Report

Quoted By:

>>7379123
>>7379117
meant for

Anonymous

Anonymous Tue 16 May 2023 21:41:36 No.7379130 Report

Quoted By:

>>7379074
>>7379085
thanks anons

Anonymous

View Same Google ImgOps iqdb SauceNAO 1661925839480185.jpg, 372KiB, 1408x1792

Anonymous Tue 16 May 2023 21:44:12 No.7379134 Report

Quoted By: >>7379202

>Make some LoHA's or whatever new toggle lora on ez scripts
>The thing feels undercooked at 4k steps
Am I missing something with this, I get character okay enough early but it cannot get hair color even close to consistent.

Anonymous

Anonymous Tue 16 May 2023 21:44:34 No.7379135 Report

Quoted By: >>7379139

Alright I'm starting off a fresh windows, what do I install to get webui up and running, and the easy lora scripts working?
>git (latest version?)
>python 3.10 (latest version? this lora guide I'm looking at says it wants 3.10.6?)
>git bash the webui
>use installer file from easy scripts git page
Is it just that simple? I assume last time I had a wrong version of python or installed more than one version. Do I manually install stuff like xformers or torch, or update pip? No idea what any of those are, but I see them in the cmd prompts a lot.

Anonymous

Anonymous Tue 16 May 2023 21:45:08 No.7379137 Report

Quoted By: >>7379146

>>7379123
I will do this since you provided catbox, I'll also use wildcards for more variety. Thanks anon

Anonymous

Anonymous Tue 16 May 2023 21:46:13 No.7379139 Report

Quoted By:

>>7379135
If you have a newer (I think it's 2000 or 3000 series and beyond) you can add --xformers to the end of set COMMANDLINE_ARGS= in the launch .bat and it will automatically install it for you when you run the application

Anonymous

View Same Google ImgOps iqdb SauceNAO 00741.png, 278KiB, 448x448

Anonymous Tue 16 May 2023 21:51:12 No.7379145 Report

Quoted By: >>7379183

Anonymous

Anonymous Tue 16 May 2023 21:51:38 No.7379146 Report

Quoted By:

>>7379137
Cool. THe prompt is a WIP so it'd be interesting to see what a large samplesize gets.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00417-3539605979-detailed pupil,.png, 2MiB, 1024x1536

Anonymous Tue 16 May 2023 21:56:34 No.7379152 Report

Quoted By: >>7384588

Anonymous

View Same Google ImgOps iqdb SauceNAO 00135-3660027748.png, 2MiB, 1024x1536

Anonymous Tue 16 May 2023 21:57:35 No.7379153 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO 00231-1755260546.png, 2MiB, 1024x1024

Anonymous Tue 16 May 2023 21:57:41 No.7379155 Report

Quoted By: >>7379172 >>7379226

Anonymous

View Same Google ImgOps iqdb SauceNAO 00497-1602371676.jpg, 930KiB, 4100x3514

Anonymous Tue 16 May 2023 22:04:49 No.7379166 Report

Quoted By: >>7379169 >>7379173 >>7379174

Do you guys notice any difference? Testing hi-res steps, and I feel above 10 it is pretty similar.

10 steps take 4s, 15 steps take 7s and 30 steps take 14s.

https://files.catbox.moe/bflwpr.png

Anonymous

Anonymous Tue 16 May 2023 22:07:11 No.7379169 Report

Quoted By: >>7379174

>>7379166
I never use them, the times I tested noticed no differences

Anonymous

View Same Google ImgOps iqdb SauceNAO 00004-3585914446.png, 3MiB, 1360x1920

Anonymous Tue 16 May 2023 22:07:42 No.7379170 Report

Quoted By: >>7379418 >>7380439

Now I'm wondering if (torch 2.1.0) cu118 caused my concept loras to not learn anything for a month+. Nightly cu117 from today seems fine but I'll have to see tomorrow with some overnight rebakes. It's really fucky since style loras got trained just fine with it.

Anonymous

Anonymous Tue 16 May 2023 22:08:07 No.7379172 Report

Quoted By:

>>7379155
suckable udder.

Anonymous

Anonymous Tue 16 May 2023 22:08:08 No.7379173 Report

Quoted By: >>7379174

>>7379166
Do you actually have the step setting turned on though?

Anonymous

Anonymous Tue 16 May 2023 22:09:07 No.7379174 Report

Quoted By:

>>7379166
Nope, hires fix steps not having any effect past a fairly low value is known

>>7379169
>>7379173
Even if you don't make use of it you should still set the slider since leaving it at 0 will use the same amount of steps as your regular sampling steps which, as you can see, is pretty redundant.

Anonymous

View Same Google ImgOps iqdb SauceNAO 11999-1321149868.png, 770KiB, 728x968

Anonymous Tue 16 May 2023 22:09:37 No.7379175 Report

Quoted By: >>7379179

Anonymous

View Same Google ImgOps iqdb SauceNAO 1683502564087479.jpg, 663KiB, 1024x1536

Anonymous Tue 16 May 2023 22:09:53 No.7379177 Report

Quoted By:

Anonymous

Anonymous Tue 16 May 2023 22:11:49 No.7379179 Report

Quoted By: >>7379194

>>7379175
Cute

Anonymous

Anonymous Tue 16 May 2023 22:13:29 No.7379182 Report

Quoted By:

>>7379025
yjk

Anonymous

Anonymous Tue 16 May 2023 22:16:20 No.7379183 Report

Quoted By:

>>7379145
based pixel art enjoyer

Anonymous

View Same Google ImgOps iqdb SauceNAO 00060-2977165292.png, 1MiB, 960x1280

Anonymous Tue 16 May 2023 22:16:52 No.7379186 Report

Quoted By:

big witch

Anonymous

View Same Google ImgOps iqdb SauceNAO 00024-344910431-1girl, hair orna (...).png, 2MiB, 1280x1792

Anonymous Tue 16 May 2023 22:21:56 No.7379190 Report

Quoted By: >>7379206

>>7378648
>what's her noble phantasm?
Pink Mitt of Infinite Cuteness

Anonymous

View Same Google ImgOps iqdb SauceNAO 00109-2699641931.png, 1019KiB, 784x1008

Anonymous Tue 16 May 2023 22:23:59 No.7379194 Report

Quoted By: >>7379195

>>7379179
Thanks

Anonymous

Anonymous Tue 16 May 2023 22:24:22 No.7379195 Report

Quoted By:

>>7379194
also cute

Anonymous

Anonymous Tue 16 May 2023 22:27:11 No.7379197 Report

Quoted By: >>7379198 >>7379210 >>7379213 >>7382066

Is xformers worth using? It's supposed to speed up generation but apparently has issues with unchanged parameters producing different results unless I'm misunderstanding.

Anonymous

Anonymous Tue 16 May 2023 22:28:13 No.7379198 Report

Quoted By: >>7379210

>>7379197
Sounds like bullshit to me t b h, there shouldn't be any downsides to using it

Anonymous

Anonymous Tue 16 May 2023 22:30:26 No.7379200 Report

Quoted By: >>7379208

>>7379036
Link other threads, copy quickstart for nvidia/amd etc, even if model advice is outdated its better than nothing.
Obviously a link to civit and huggingface.
new rentry or OP mini-category, call it prompt help or whatever, take artistssd and faces-faces-faces link, from https://rentry.org/sdg-link#prompt-sharing and the 3 links to chinese docs
https://rentry.org/59xed3 add this under a "LoRA training" in OP and move easyscripts there

Anonymous

Anonymous Tue 16 May 2023 22:32:01 No.7379202 Report

Quoted By:

>>7379134
>falling for retarded new thing

Anonymous

Anonymous Tue 16 May 2023 22:33:53 No.7379206 Report

Quoted By: >>7379246

>>7379190
more like Pink Mitt of Disability or Pink Mitt of Mangled Hand. Imparts a horrible curse if it touches you.

Anonymous

Anonymous Tue 16 May 2023 22:35:22 No.7379208 Report

Quoted By: >>7379217

>>7379200
>DO NOT INSTALL IT IN THE SAME PLACE AS YOUR WEBUI INSTALL! MAKE A NEW PYTHON VIRTUAL ENV!
What does this mean? Just git bash in a different folder? Or do I have to google python virtual environments?

Also, do I have to install kohya scripts or can I just use the lora easy scripts installer?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00523-2275262505.png, 2MiB, 1280x1280

Anonymous Tue 16 May 2023 22:35:39 No.7379209 Report

Quoted By: >>7379214

Does someone knows if using the "<lora:a:1>" on the prompt, is affected if you use it at the start or at the end of the prompt?

For example, is there a diff with "<lora:a:1>, prompt" vs "prompt, <lora:a:1>"?

Anonymous

Anonymous Tue 16 May 2023 22:35:57 No.7379210 Report

Quoted By: >>7379213

>>7379198
newfag

>>7379197
Just test it yourself, its very minor differences on the same seed - there's no real reason to ever regenerate an image 1:1 in the first place so its not exactly a drawback. Best case it might fix a finger if you reroll enough times but its still faster to inpaint - its entirely inconsequential.

Anonymous

View Same Google ImgOps iqdb SauceNAO Synchronicity.jpg, 239KiB, 1024x1536

Anonymous Tue 16 May 2023 22:36:37 No.7379212 Report

Quoted By: >>7379216 >>7379229

Anonymous

View Same Google ImgOps iqdb SauceNAO 00524-2619683666.png, 2MiB, 1280x1280

Anonymous Tue 16 May 2023 22:37:00 No.7379213 Report

Quoted By: >>7379217 >>7379263 >>7379267

>>7379197
>>7379197
Since xformers 0.0.20, it is determistic so there's no downside now. Well, the downside it may be that you have to build it

https://github.com/AUTOMATIC1111/stable-diffusion-webui/discussions/2103

I built it the other day for torch+cu121

>>7379210
Since latest torchs, it is determistic so there's not even the downside of difference when regenerating an image

Anonymous

Anonymous Tue 16 May 2023 22:37:41 No.7379214 Report

Quoted By:

>>7379209
there is no difference

Anonymous

Anonymous Tue 16 May 2023 22:39:13 No.7379216 Report

Quoted By: >>7379229

>>7379212
holy
catbox, workflow?

Anonymous

Anonymous Tue 16 May 2023 22:39:42 No.7379217 Report

Quoted By: >>7379221

>>7379208
>schizo babble
kill yourself retard, I have no clue what you're talking about

>>7379213
>have to build it
yeah no. Is it even faster or eats less VRAM at least?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00529-1578477623.png, 2MiB, 1280x1280

Anonymous Tue 16 May 2023 22:41:49 No.7379221 Report

Quoted By: >>7379223

>>7379217
>yeah no. Is it even faster or eats less VRAM at least?
It is faster, since it is built specially for your gpu architecture vs official wheels. VRAM seems to be the same, or at least I haven't noticed a difference there.

Anonymous

Anonymous Tue 16 May 2023 22:42:25 No.7379223 Report

Quoted By: >>7379228

>>7379221
how much faster

Anonymous

Anonymous Tue 16 May 2023 22:45:12 No.7379226 Report

Quoted By:

>>7379155
Really nice, mind posting a box? Love the asymmetrical dress
Also did you use the nyx lora from civ, since the example images look pretty underwhelming?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00531-4218347079.png, 2MiB, 1280x1280

Anonymous Tue 16 May 2023 22:45:45 No.7379228 Report

Quoted By: >>7379232

>>7379223
5% on inference and prob 10% on training. I can do some tests later, since I would have to uninstall xformers/torch I'm using to install torch 2.0.0+cu118 and xformers 0.0.17.

(I'm on torch 2.1.0+cu121)

It is "maybe" more noticeable while training LoRAs.

But more than the speed, getting the determistic results is 10/10.

Anonymous

View Same Google ImgOps iqdb SauceNAO Synchronicity.webm, 2MiB, 512x768

Anonymous Tue 16 May 2023 22:46:49 No.7379229 Report

Quoted By: >>7379233 >>7379238 >>7380023

>>7379212
>>7379216

Anonymous

Anonymous Tue 16 May 2023 22:48:44 No.7379232 Report

Quoted By:

>>7379228
>getting the determistic results is 10/10.
I never got why anyone gives a shit about this. What is the point of making the exact same image as the one you already have? What am I not seeing for this to be "10/10"?

Anonymous

Anonymous Tue 16 May 2023 22:48:55 No.7379233 Report

Quoted By: >>7379235

>>7379229
this is some kinda friggin cope around learning to draw

Anonymous

Anonymous Tue 16 May 2023 22:48:57 No.7379234 Report

Quoted By: >>7379243 >>7379247 >>7379248

Is regional prompter meant to feel like it works 1% of the time

Anonymous

Anonymous Tue 16 May 2023 22:50:42 No.7379235 Report

Quoted By: >>7379914 >>7379939 >>7380023

>>7379233
this took about 1h20 end to end, from the first gen to the final image. about 127 inpaints.
do you think someone can produce a similar result that quickly?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00143-2031944928.png, 986KiB, 896x896

Anonymous Tue 16 May 2023 22:51:19 No.7379237 Report

Quoted By:

Anonymous

Anonymous Tue 16 May 2023 22:51:42 No.7379238 Report

Quoted By:

>>7379229
Ah so you really just went hard with the editing/inpainting.
Well done on that one.

Anonymous

View Same Google ImgOps iqdb SauceNAO movida 2.png, 41KiB, 1280x720

Anonymous Tue 16 May 2023 22:51:49 No.7379239 Report

Quoted By: >>7379242 >>7379249 >>7379252 >>7379257 >>7382471

It's hilarious to me to go from this..

Anonymous

View Same Google ImgOps iqdb SauceNAO anime__dark_hair_man__erected_di (...).jpg, 137KiB, 1024x768

Anonymous Tue 16 May 2023 22:52:47 No.7379242 Report

Quoted By: >>7379249 >>7379252

>>7379239
to this with just two steps in the middle.

Anonymous

Anonymous Tue 16 May 2023 22:52:58 No.7379243 Report

Quoted By: >>7379247

>>7379234
It was pretty janky the last time I tried using it, no idea how it is after the updates that have come since then. I don't really do prompts that'd benefit from using it but the last time I was playing around with the multidiffusion upscaler I found that it worked better than regional prompter
Even if regional prompter somehow is better than multidiffusion upscaler (I don't know if it is) it's still worth it to get the latter just for the img2img upscaler

Anonymous

View Same Google ImgOps iqdb SauceNAO 00025-4089271057-1girl, hair orn (...).png, 2MiB, 1280x1792

Anonymous Tue 16 May 2023 22:54:13 No.7379246 Report

Quoted By: >>7379255

>>7379206
She does part-time at a cursed mansion too like that other guy.

Anonymous

Anonymous Tue 16 May 2023 22:54:21 No.7379247 Report

Quoted By:

>>7379234
no. worked fine for me.

>>7379243
multidiffusion is shit, its useful for foreground imitations of in-game sprites and thats about it. The seams are too obvious

Anonymous

Anonymous Tue 16 May 2023 22:54:23 No.7379248 Report

Quoted By: >>7379254

>>7379234
This but latent couple

Anonymous

Anonymous Tue 16 May 2023 22:54:38 No.7379249 Report

Quoted By:

>>7379239
>>7379242
based

Anonymous

Anonymous Tue 16 May 2023 22:55:23 No.7379252 Report

Quoted By:

>>7379239
sovl
>>7379242
sovlless

Anonymous

Anonymous Tue 16 May 2023 22:56:00 No.7379254 Report

Quoted By: >>7379259 >>7379261

>>7379248
Latent couple hasn't been updated in months, it's completely abandoned. One of the reasons outdated shit should just get removed from the OP

Anonymous

Anonymous Tue 16 May 2023 22:56:20 No.7379255 Report

Quoted By:

>>7379246
Glad you thought of it too, couldnt think of a missing link between the two.

Anonymous

Anonymous Tue 16 May 2023 22:58:02 No.7379257 Report

Quoted By: >>7379269

>>7379239
lol what did you do for this? just controlnet?

Anonymous

Anonymous Tue 16 May 2023 22:59:03 No.7379258 Report

Quoted By:

BREAK's pretty useful for color contamination.

Anonymous

Anonymous Tue 16 May 2023 23:00:23 No.7379259 Report

Quoted By:

>>7379254
Well fuck me

Anonymous

Anonymous Tue 16 May 2023 23:01:24 No.7379261 Report

Quoted By:

>>7379254
>hasn't been updated in months
oh so it still works then

Anonymous

Anonymous Tue 16 May 2023 23:05:50 No.7379263 Report

Quoted By: >>7379273

>>7379213
>0.0.20
0.0.19 is deterministic and has prebuilt wheels on pip (just pip install)

Anonymous

Anonymous Tue 16 May 2023 23:08:06 No.7379267 Report

Quoted By: >>7379273

>>7379213
well that guide didn't work at all

Anonymous

View Same Google ImgOps iqdb SauceNAO 1621245534531.png, 480KiB, 512x768

Anonymous Tue 16 May 2023 23:09:41 No.7379269 Report

Quoted By:

>>7379257
not even, just easy diffusion, anyhentai as a model and lets go.

Now I'm messing with known images to H. but they still don't go full anime.

Anonymous

View Same Google ImgOps iqdb SauceNAO wat.png, 1MiB, 784x1008

Anonymous Tue 16 May 2023 23:12:10 No.7379271 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO 00538-630400567.png, 2MiB, 1280x1536

Anonymous Tue 16 May 2023 23:12:23 No.7379273 Report

Quoted By: >>7379279 >>7379295

>>7379263
Oh naisu, that way it's just install and play

>>7379267
Really? I did build on Sunday and had no issues. You have on step 4, install the torch you want to use (stable is torch2.0.1+cu118 I think), and then skip step 6, since that was when building torch+cu113

Anonymous

Anonymous Tue 16 May 2023 23:16:58 No.7379279 Report

Quoted By: >>7379283 >>7379294

>>7379273
xformers 0.0.19 requires torch==2.0.0, torchvision==0.15.1, so blindly updating will fuck your shit up.

Anonymous

Anonymous Tue 16 May 2023 23:18:12 No.7379281 Report

Quoted By: >>7379397

Been away from prompting for a while.
HD22 gives pale pictures for some reason and novelAI v3 never seems to produce penetration.
What's a to go model that doesn't require fiddling with Loras? I've seen another anon ask and am looking into it, but what's like the most generic thing?

Anonymous

Anonymous Tue 16 May 2023 23:18:45 No.7379283 Report

Quoted By: >>7379286

>>7379279
don't care doing it anyway

Anonymous

Anonymous Tue 16 May 2023 23:19:17 No.7379285 Report

Quoted By: >>7379299

>/trash/ sdg actually putting out better images than us
it's over...

Anonymous

Anonymous Tue 16 May 2023 23:19:21 No.7379286 Report

Quoted By:

>>7379283
based retard

Anonymous

View Same Google ImgOps iqdb SauceNAO 00335-488343187.png, 2MiB, 1024x1536

Anonymous Tue 16 May 2023 23:21:48 No.7379290 Report

Quoted By: >>7379301

Anonymous

Anonymous Tue 16 May 2023 23:29:14 No.7379294 Report

Quoted By: >>7379417

>>7379279 (me)
Oh, looks like xformers 0.0.20 has been prereleased so you should be able to just do
pip install -U --pre xformers torch torchvision

Anonymous

Anonymous Tue 16 May 2023 23:29:17 No.7379295 Report

Quoted By: >>7379305

>>7379273
I'm retarded, how do I check what version of torch I have in webui?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00548-3975800048.png, 2MiB, 1160x1696

Anonymous Tue 16 May 2023 23:29:17 No.7379296 Report

Quoted By: >>7379578

idk what happened here, but i'm scared

Anonymous

View Same Google ImgOps iqdb SauceNAO 00001-3015775106.png, 2MiB, 1024x1536

Anonymous Tue 16 May 2023 23:30:15 No.7379299 Report

Quoted By: >>7379302

>>7379285
>surprised at the technical prowess of furry coomers
filter the avatarfags and their blog posts, the people that think using a lora is too advanced, and the fully clothed 1girl /g/ tier posts, then you'll see that the actual on-topic posts are breddy gud

Anonymous

View Same Google ImgOps iqdb SauceNAO 00027-2467287746.png, 2MiB, 1024x1536

Anonymous Tue 16 May 2023 23:30:38 No.7379300 Report

Quoted By: >>7380356

Auto tagger + img2img to convert anything to anime is pretty fun.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00320-3147540639-nsfw, scan, (tr (...).png, 2MiB, 1024x1536

Anonymous Tue 16 May 2023 23:32:28 No.7379301 Report

Quoted By:

>>7379290
hot

Anonymous

Anonymous Tue 16 May 2023 23:33:21 No.7379302 Report

Quoted By:

>>7379299
Seeing an actual polar bear suck off a dude right off the bat convinced me they're gods.

Anonymous

Anonymous Tue 16 May 2023 23:35:19 No.7379305 Report

Quoted By:

>>7379295
Its on the bottom of the interface, which is where you can also reload UI without going to the settings

Anonymous

View Same Google ImgOps iqdb SauceNAO digits.png, 1MiB, 872x992

Anonymous Tue 16 May 2023 23:39:54 No.7379311 Report

Quoted By: >>7379312

Anonymous

Anonymous Tue 16 May 2023 23:41:03 No.7379312 Report

Quoted By:

>>7379311
based

Anonymous

Anonymous Tue 16 May 2023 23:42:55 No.7379313 Report

Quoted By: >>7379317 >>7379353

>tfw racism strikes again
AI cant figure out afros

Anonymous

View Same Google ImgOps iqdb SauceNAO 00470-597925038-best quality, si (...).png, 805KiB, 768x1120

Anonymous Tue 16 May 2023 23:44:53 No.7379314 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO 00000-757670233-1girl, brown ski (...).png, 745KiB, 640x896

Anonymous Tue 16 May 2023 23:46:05 No.7379317 Report

Quoted By: >>7379339

>>7379313
hm?
But true, it struggles if you use loras for some reason.

Anonymous

Anonymous Tue 16 May 2023 23:49:50 No.7379319 Report

Quoted By: >>7379320 >>7379343

am I right in that a quick and easy way to tell if a lora is overbaked is to look for features the training data had but the prompt does not?
so if my character had a hat in more than a few, but I didn't tell it to add one, it shouldn't be adding one?

Anonymous

Anonymous Tue 16 May 2023 23:51:19 No.7379320 Report

Quoted By: >>7379331

>>7379319
Something like that. More obvious if you negative a hat and it still does it.
But I like to test sex poses and specific clothing and seeing what it does more.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00442-1771704806-detailed pupil,.png, 2MiB, 1024x1536

Anonymous Tue 16 May 2023 23:54:59 No.7379323 Report

Quoted By: >>7379422

Anonymous

View Same Google ImgOps iqdb SauceNAO 3steps.png, 1MiB, 1024x1024

Anonymous Tue 16 May 2023 23:57:49 No.7379330 Report

Quoted By:

Another one, It's so funny. from the first image, to the second, and from the second to the other two, (the last one I found it extra funny for the change in perspective)
and this is like 15 minutes while I play wildfrost.

Anonymous

Anonymous Tue 16 May 2023 23:57:52 No.7379331 Report

Quoted By: >>7379338 >>7379604

>>7379320
a shame this kind of backfires
going far back enough that the overbaking goes away is far back enough for it to "forget" what the correct headwear looks like, almost as if I need to make another fucking lora just for it

Anonymous

Anonymous Wed 17 May 2023 00:01:57 No.7379338 Report

Quoted By: >>7379357

>>7379331
Try a resize first.

Anonymous

Anonymous Wed 17 May 2023 00:02:15 No.7379339 Report

Quoted By: >>7379341 >>7379346

>>7379317
yeah, afros are harder than they have any right to be.

Anonymous

Anonymous Wed 17 May 2023 00:04:06 No.7379341 Report

Quoted By: >>7379346

>>7379339
Well, japs hate blacks AND curly hair. It was rigged from the start.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00566-818752806.jpg, 767KiB, 4100x2530

Anonymous Wed 17 May 2023 00:05:05 No.7379343 Report

Quoted By:

>>7379319
>so if my character had a hat in more than a few, but I didn't tell it to add one, it shouldn't be adding one?

Not sure how it goes there, but for example, I traned a lot of character LoRAs, but for this case I trained Mona from Genshin Impact. My dataset probably had images with hats 50% of the time.

When I prompt the character with the lora and just activation token, it doesn't add the hat, but when I prompt the hat it adds it.

I didn't remove the hat tag from the dataset.

Here is an example image.

Anonymous

Anonymous Wed 17 May 2023 00:05:41 No.7379345 Report

Quoted By:

interdasting, I think all I managed to do was change from accelerator v12 to v18 and it sped me up a shiteload

Anonymous

Anonymous Wed 17 May 2023 00:05:41 No.7379346 Report

Quoted By:

>>7379339
>>7379341 me
The one I always weep for is the kino short sides curly hair.
Damn near impossible to prompt well without seventy emphasises and retries.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00377-2401094739.png, 1MiB, 1536x1024

Anonymous Wed 17 May 2023 00:07:25 No.7379351 Report

Quoted By: >>7388582

Anonymous

View Same Google ImgOps iqdb SauceNAO 22094-670504926-1girl, nigger, a (...).png, 1MiB, 1024x1024

Anonymous Wed 17 May 2023 00:10:46 No.7379353 Report

Quoted By:

>>7379313
it recognizes the n word, somehow. see if that would help afros

>attempt 1

Anonymous

View Same Google ImgOps iqdb SauceNAO __belina_dappou_tamer_no_nariaga (...).jpg, 562KiB, 796x2000

Anonymous Wed 17 May 2023 00:10:47 No.7379354 Report

Quoted By: >>7379940

Anyone done an ookuma nekosuke LORA?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00054-3501946613-1girl, two side (...).png, 1MiB, 640x896

Anonymous Wed 17 May 2023 00:15:37 No.7379356 Report

Quoted By: >>7379382

hmmm

Anonymous

Anonymous Wed 17 May 2023 00:17:16 No.7379357 Report

Quoted By: >>7379360

>>7379338
I made it at 128, is 64 a good resize?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00074-396723298-1girl, two side (...).png, 1023KiB, 640x896

Anonymous Wed 17 May 2023 00:19:17 No.7379360 Report

Quoted By:

Apparently ancient reliefs need a lora.
>>7379357
I do 64 and 32 and compare.
Under 32 it starts getting diluted.

Anonymous

View Same Google ImgOps iqdb SauceNAO 20230512_212752.jpg, 105KiB, 444x691

Anonymous Wed 17 May 2023 00:22:19 No.7379364 Report

Quoted By: >>7379368 >>7379382 >>7379399

I wanma make something like this. Is there a way to limit the color used?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00297-2656964295.png, 926KiB, 896x896

Anonymous Wed 17 May 2023 00:23:35 No.7379367 Report

Quoted By:

Anonymous

Anonymous Wed 17 May 2023 00:23:50 No.7379368 Report

Quoted By:

>>7379364
I believe someone asked the same thing last thread.

Anonymous

Anonymous Wed 17 May 2023 00:25:06 No.7379369 Report

Quoted By:

Is there a way to scrape top Google/search engine results using Grabber or similar?

Anonymous

Anonymous Wed 17 May 2023 00:27:09 No.7379373 Report

Quoted By:

Has anyone ever managed a lora for wings?
any time I've tried its been an absolute fucking write off

Anonymous

View Same Google ImgOps iqdb SauceNAO 00262-931096305.png, 2MiB, 1024x1536

Anonymous Wed 17 May 2023 00:29:59 No.7379376 Report

Quoted By: >>7379576

so retrained thing in 1024 size images , im unsure if its worth the effort.

Anonymous

Anonymous Wed 17 May 2023 00:31:04 No.7379380 Report

Quoted By: >>7379384

https://arxiv.org/abs/2301.00774

Anonymous

Anonymous Wed 17 May 2023 00:32:20 No.7379382 Report

Quoted By: >>7379409

>>7379356
kek

>>7379364
try monochrome teal and monochrome pink.

Anonymous

Anonymous Wed 17 May 2023 00:34:14 No.7379384 Report

Quoted By: >>7379387 >>7379388

>>7379380
>/arxiv.org/
im not clicking that. ,l/

Anonymous

View Same Google ImgOps iqdb SauceNAO 99312-2699621930.png, 889KiB, 784x1008

Anonymous Wed 17 May 2023 00:36:05 No.7379386 Report

Quoted By:

Anonymous

Anonymous Wed 17 May 2023 00:39:44 No.7379387 Report

Quoted By:

>>7379384
>he doesn't know what arxiv is
just put "Retard" in the name field next time

Anonymous

Anonymous Wed 17 May 2023 00:41:27 No.7379388 Report

Quoted By:

>>7379384
You can look up what it is on bing/jewgle/wherever
its a fucking research paper archive

Anonymous

View Same Google ImgOps iqdb SauceNAO 01919-1534093253.png, 2MiB, 1536x1024

Anonymous Wed 17 May 2023 00:49:17 No.7379397 Report

Quoted By: >>7379839 >>7379999 >>7380183

>>7379281
>HD22 gives pale pictures for some reason
You need to load a VAE. This one is fairly popular:
https://huggingface.co/Yukihime256/840000/tree/main

>What's a to go model that doesn't require fiddling with Loras? I've seen another anon ask and am looking into it, but what's like the most generic thing?
AbyssOrangeMix2_hard, Butter, or Based65 (good if you don't want to use any LoRAs, otherwise use Based64).

Anonymous

Anonymous Wed 17 May 2023 00:50:01 No.7379399 Report

Quoted By: >>7379476

>>7379364
flat colors, limited palette, watercolor, physical media

Anonymous

View Same Google ImgOps iqdb SauceNAO 02012-3886179060.png, 2MiB, 1536x1024

Anonymous Wed 17 May 2023 00:56:03 No.7379409 Report

Quoted By: >>7379421

>>7379382
>try monochrome teal and monochrome pink.
"pink theme" is a known Danbooru tag. "teal theme" isn't, but NAI-based models might still be able to recognize it. This pic wasn't even intended to be monochrome, but "garishly colored alien lichens" tends to override either the character's hair color or the entire image.

Anonymous

Anonymous Wed 17 May 2023 01:02:54 No.7379413 Report

Quoted By:

it's over

Anonymous

Anonymous Wed 17 May 2023 01:04:54 No.7379417 Report

Quoted By:

>>7379294
Is there any issues for using the pre release 0,0.20 over the stable 0.0.19 xformers

Anonymous

Anonymous Wed 17 May 2023 01:05:10 No.7379418 Report

Quoted By:

>>7379170
i've noticed big differences with training on different torch versions. do post an update on this because i am also curious

Anonymous

Anonymous Wed 17 May 2023 01:06:12 No.7379421 Report

Quoted By:

>>7379409
That's a good suggestion. I'd go with yours over mine.

Anonymous

Anonymous Wed 17 May 2023 01:07:37 No.7379422 Report

Quoted By: >>7380074

>>7379323
catbox for this one?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00150-146740117.png, 2MiB, 1024x1536

Anonymous Wed 17 May 2023 01:12:03 No.7379425 Report

Quoted By: >>7379429 >>7379458 >>7379470

>rebaked to fuck around with shuffle captions gacha
>truck kun randomly appears
kek

Anonymous

Anonymous Wed 17 May 2023 01:12:37 No.7379426 Report

Quoted By: >>7379428

doctors says I need to fap less

Anonymous

Anonymous Wed 17 May 2023 01:21:33 No.7379428 Report

Quoted By: >>7379430

>>7379426
I wonder if urologists worldwide are swarmed with bruised dicks
Even back when nai just leaked there were japs blogging on twitter how they had to get ointment for their demolished cocks

Anonymous

Anonymous Wed 17 May 2023 01:22:50 No.7379429 Report

Quoted By:

>>7379425
oh no she's about to get isekai'd

Anonymous

Anonymous Wed 17 May 2023 01:24:59 No.7379430 Report

Quoted By: >>7379891

>>7379428
>bruised dicks
no , I have an enlarged prostate which makes it hard for me to pee, psa test came back suggesting its not cancer but it may just be inflamed from overuse and I need to take it easy.

Anonymous

Anonymous Wed 17 May 2023 01:26:24 No.7379432 Report

Quoted By: >>7379435 >>7379459

Are there any new block merges or finetunes? Or is it all LoRA-merged fried models now?

Anonymous

Anonymous Wed 17 May 2023 01:27:10 No.7379435 Report

Quoted By:

>>7379432
>>7379432
>Or is it all LoRA-merged fried models now?
This is the first thing I remember
Now it's the last thing left on my mind

Anonymous

Anonymous Wed 17 May 2023 01:39:45 No.7379444 Report

Quoted By: >>7379447 >>7379456

Can someone recommend a website other than Collab to try Stable Diffussion

Anonymous

View Same Google ImgOps iqdb SauceNAO 00065-2562761652.png, 2MiB, 744x1264

Anonymous Wed 17 May 2023 01:42:16 No.7379446 Report

Quoted By:

does anyone know how i can get the botw style lora to work with a midna lora?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00357-3371433151.png, 1MiB, 1536x1024

Anonymous Wed 17 May 2023 01:42:48 No.7379447 Report

Quoted By: >>7379454

>>7379444
That reminds me that the OP should tell people that free colabs are kill.

Anonymous

Anonymous Wed 17 May 2023 01:44:09 No.7379449 Report

Quoted By: >>7379450 >>7379452

>>7379025
are fingers good now?

Anonymous

Anonymous Wed 17 May 2023 01:45:06 No.7379450 Report

Quoted By:

>>7379449
not often

Anonymous

Anonymous Wed 17 May 2023 01:49:09 No.7379452 Report

Quoted By:

>>7379449
not without inpainting or controlnet

Anonymous

Anonymous Wed 17 May 2023 01:50:55 No.7379454 Report

Quoted By:

>>7379447
More like they realised how rentable it is and now are behind paywalls

Anonymous

Anonymous Wed 17 May 2023 01:53:47 No.7379456 Report

Quoted By:

>>7379444
>try
https://www.youtube.com/watch?v=BQ4yd2W50No

Anonymous

Anonymous Wed 17 May 2023 01:55:55 No.7379458 Report

Quoted By:

>>7379425
zamn that's a nice truck actually

Anonymous

Anonymous Wed 17 May 2023 01:57:57 No.7379459 Report

Quoted By: >>7379469

>>7379432
the latter. Get used to animelike2D ad infinitum

Anonymous

View Same Google ImgOps iqdb SauceNAO nazielf.png, 1MiB, 1024x1024

Anonymous Wed 17 May 2023 01:58:30 No.7379460 Report

Quoted By: >>7379471 >>7380489

Anonymous

Anonymous Wed 17 May 2023 01:59:06 No.7379461 Report

Quoted By: >>7379462 >>7379469

>new 4090 only gets 1.3it/s
what'd I fuck up this time?
I'm on torch 2.0, cuda118

Anonymous

Anonymous Wed 17 May 2023 01:59:45 No.7379462 Report

Quoted By: >>7379464 >>7379597

>>7379461
genning on cpu?

Anonymous

Anonymous Wed 17 May 2023 02:01:08 No.7379464 Report

Quoted By:

>>7379462
Task manager says no, CPU is not under load while the card is

Anonymous

Anonymous Wed 17 May 2023 02:05:50 No.7379469 Report

Quoted By: >>7379473

>>7379461
probably missing xformers or sdp-no-mem-attention
>>7379459
the merge with model option should never have been added, it's over

Anonymous

Anonymous Wed 17 May 2023 02:05:52 No.7379470 Report

Quoted By: >>7379479 >>7380040

>>7379425
i really like the style on this one. catbox?

Anonymous

View Same Google ImgOps iqdb SauceNAO catbox_ibioxq.png, 1MiB, 1024x1024

Anonymous Wed 17 May 2023 02:06:04 No.7379471 Report

Quoted By:

>>7379460

Anonymous

View Same Google ImgOps iqdb SauceNAO 00017-165063295.png, 487KiB, 512x768

Anonymous Wed 17 May 2023 02:07:05 No.7379472 Report

Quoted By:

Anonymous

Anonymous Wed 17 May 2023 02:07:10 No.7379473 Report

Quoted By:

>>7379469
>xformers
I guess I'm retarded since I was told not to bother with that
weirdly enough, I do have spd-no-mem-attention and it seems to have made the problem worse, since I was getting better rates from my 3070, I didn't have xformers then either

Anonymous

Anonymous Wed 17 May 2023 02:08:42 No.7379476 Report

Quoted By:

>>7379399
this. limited palette helps a lot, but sure as fuck aint perfect.

Anonymous

Anonymous Wed 17 May 2023 02:12:11 No.7379479 Report

Quoted By: >>7379480 >>7380040

>>7379470
unreleased lora i'll probably finish tomorrow morning

Anonymous

Anonymous Wed 17 May 2023 02:14:38 No.7379480 Report

Quoted By:

>>7379479
nice. looking really good so far.

Anonymous

View Same Google ImgOps iqdb SauceNAO catbox_4t0ito.png, 1MiB, 1024x1024

Anonymous Wed 17 May 2023 02:24:18 No.7379484 Report

Quoted By: >>7379508

Anonymous

View Same Google ImgOps iqdb SauceNAO 00801-3118678930.jpg, 298KiB, 1024x1536

Anonymous Wed 17 May 2023 02:33:53 No.7379496 Report

Quoted By: >>7379498 >>7379505

>last night you passed out while doing a ritual to summon a demon
>wake up next day
>you can hear someone in your house
>go investigate
>find pic related in corridor
wutdo?

Anonymous

Anonymous Wed 17 May 2023 02:35:23 No.7379498 Report

Quoted By:

>>7379496
>wutdo?
get banned I guess lmao

Anonymous

View Same Google ImgOps iqdb SauceNAO catbox_2ruwfq.png, 1MiB, 1024x1024

Anonymous Wed 17 May 2023 02:35:25 No.7379499 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO 00134-170316253.png, 2MiB, 1024x1536

Anonymous Wed 17 May 2023 02:38:01 No.7379505 Report

Quoted By: >>7379514 >>7381552 >>7381554

>>7379496
put your trip ban so I can filter you, I'm not getting a warning over a report again

Anonymous

Anonymous Wed 17 May 2023 02:38:41 No.7379508 Report

Quoted By:

>>7379484
damn she's a big girl.

Anonymous

View Same Google ImgOps iqdb SauceNAO catbox_4wwe88.png, 1MiB, 1024x1024

Anonymous Wed 17 May 2023 02:40:02 No.7379512 Report

Quoted By:

tfw youll never go on a world tour of all the elf kingdoms blowing fat loads in every type of elf

Anonymous

Anonymous Wed 17 May 2023 02:42:02 No.7379514 Report

Quoted By: >>7379516 >>7379521

>>7379505
>reporting a drawing made by a computer that clearly looks like a drawing/3dmodel

You can only blame yourself, anon.

Anonymous

Anonymous Wed 17 May 2023 02:43:02 No.7379516 Report

Quoted By: >>7379599

>>7379514
loli is still against /h/ rules, regardless.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00133-170316250.png, 2MiB, 1024x1536

Anonymous Wed 17 May 2023 02:44:09 No.7379521 Report

Quoted By: >>7381029

>>7379514
>already nuked
kek get fucked.

Anonymous

View Same Google ImgOps iqdb SauceNAO catbox_8w520g.png, 1MiB, 1024x1024

Anonymous Wed 17 May 2023 02:49:39 No.7379530 Report

Quoted By: >>7379534

>High elves
>Drow
>Wood elves
>Nazi(Vampire?) elves
Am I missing any?

Anonymous

Anonymous Wed 17 May 2023 02:51:33 No.7379534 Report

Quoted By:

>>7379530
keebler

Anonymous

View Same Google ImgOps iqdb SauceNAO 00047-2617144973.png, 562KiB, 768x512

Anonymous Wed 17 May 2023 02:55:00 No.7379547 Report

Quoted By: >>7379557 >>7379559

Fffffffff
>try to have cute picture of girl petting a capybara at the zoo for the theme
>not only does she refuse to pet the capybara, she refuses to keep her top on.

Anonymous

Anonymous Wed 17 May 2023 02:56:47 No.7379549 Report

Quoted By: >>7379606

https://github.com/gradio-app/gradio/blob/main/CHANGELOG.md#3310
>Gradio will no longer send any analytics or call home if analytics are disabled with the GRADIO_ANALYTICS_ENABLED environment variable
/g/ freaked out about this months ago so I stepped in to actually help fix it. you're welcome.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00499-2704700392-masterpiece, la (...).png, 590KiB, 656x960

Anonymous Wed 17 May 2023 02:58:25 No.7379553 Report

Quoted By: >>7379824

Anonymous

Anonymous Wed 17 May 2023 03:01:05 No.7379557 Report

Quoted By:

>>7379547
ugh women am I right?

Anonymous

Anonymous Wed 17 May 2023 03:01:27 No.7379559 Report

Quoted By:

>>7379547
or her bottom from the looks of it

Anonymous

Anonymous Wed 17 May 2023 03:04:55 No.7379563 Report

Quoted By: >>7379566

>can't load resized LORA with dim from weights
why does this even exist

Anonymous

Anonymous Wed 17 May 2023 03:06:05 No.7379566 Report

Quoted By:

>>7379563
skill issue

Anonymous

Anonymous Wed 17 May 2023 03:07:03 No.7379571 Report

Quoted By:

What are people using to tag their images for training these days? wd1.4? deepdanbooru3/4? something new?

Anonymous

Anonymous Wed 17 May 2023 03:07:39 No.7379573 Report

Quoted By: >>7379593 >>7379629

imagine training and not just endlessly proompting

Anonymous

Anonymous Wed 17 May 2023 03:09:26 No.7379575 Report

Quoted By:

imagine prompting and not just endlessly training

Anonymous

Anonymous Wed 17 May 2023 03:10:28 No.7379576 Report

Quoted By:

>>7379376
Cute Sanya

Anonymous

Anonymous Wed 17 May 2023 03:11:24 No.7379577 Report

Quoted By:

imagine prompting and not just endlessly proompting

Anonymous

Anonymous Wed 17 May 2023 03:12:07 No.7379578 Report

Quoted By:

>>7379296
Too much head room

Anonymous

View Same Google ImgOps iqdb SauceNAO 108177648_p4.png, 4MiB, 1160x1728

Anonymous Wed 17 May 2023 03:12:07 No.7379579 Report

Quoted By: >>7379587 >>7379624 >>7382258

https://kix.fanbox.cc/ (Source for image)
Is there a decent way to archive fanbox content? Pixiv is going to be banning a lot of AI art (Lora resource or not) off of fanbox. Kemono does not grab anything if you aren't/weren't subbed.

Semi-related: Is there a guide/even a method for creating a lora for a certain type of clothing? (annoyed right now about how badly grass skirts and loincloths tend to come out in a lot of models).

Anonymous

Anonymous Wed 17 May 2023 03:14:15 No.7379587 Report

Quoted By:

>>7379579
gallery-dl lets you scrape it iirc

Anonymous

View Same Google ImgOps iqdb SauceNAO 00432-3283754235.png, 2MiB, 1024x1536

Anonymous Wed 17 May 2023 03:16:27 No.7379593 Report

Quoted By: >>7380277 >>7388582

>>7379573
Don't have to.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00516-1109880107-best quality, s (...).png, 549KiB, 768x768

Anonymous Wed 17 May 2023 03:17:31 No.7379596 Report

Quoted By:

Anonymous

Anonymous Wed 17 May 2023 03:17:37 No.7379597 Report

Quoted By:

>>7379462
aint no way you're getting 1.3it/s on a cpu

Anonymous

Anonymous Wed 17 May 2023 03:20:04 No.7379599 Report

Quoted By: >>7379608

>>7379516
why did loli ever get sent to /b/ anyway? seems like a weird place for it

Anonymous

Anonymous Wed 17 May 2023 03:20:07 No.7379600 Report

Quoted By: >>7379613 >>7379640

https://arxiv.org/abs/2305.08891

Diffusion generation is fundamentally flawed.

Our new paper analyzes the noise schedule and sample steps used in Stable Diffusion models and found that they are very flawed.

The existing designs cause the images to always have plain medium brightness. After fixing it, SD can generate much darker/brighter and more cinematic images.

This is different from offset noise.This addresses the issue at a more fundamental level.

Anonymous

Anonymous Wed 17 May 2023 03:21:25 No.7379604 Report

Quoted By: >>7379615

>>7379331
thats called either shit settings or bad tagging

Anonymous

Anonymous Wed 17 May 2023 03:21:55 No.7379606 Report

Quoted By: >>7379607

>>7379549
iirc it's still sending requests even with analytics disabled, the only way to prevent it is to redirect requests

Anonymous

Anonymous Wed 17 May 2023 03:22:36 No.7379607 Report

Quoted By: >>7379611

>>7379606
That was the point of the PRs made, to fix that.

Anonymous

Anonymous Wed 17 May 2023 03:23:00 No.7379608 Report

Quoted By: >>7379646

>>7379599
most people dont wanna see that shit, and it aint illegal in the us, so to /b/ it goes.

Anonymous

Anonymous Wed 17 May 2023 03:24:39 No.7379610 Report

Quoted By: >>7379623

Anything wrong with using a gui to train loras? I tried this bmaltais one and it just worked, much better than following a rentry guide.

Anonymous

Anonymous Wed 17 May 2023 03:25:15 No.7379611 Report

Quoted By:

>>7379607
I'm also illiterate, so...

Anonymous

Anonymous Wed 17 May 2023 03:25:44 No.7379613 Report

Quoted By: >>7379626

>>7379600
Was just reading this. Kinda sucks you need to rely on some group to splurge the $$$ and fix other existing models with this now.

Anonymous

Anonymous Wed 17 May 2023 03:26:03 No.7379615 Report

Quoted By: >>7379630

>>7379604
got settings recommendations?
The lora works fine everywhere else, looking deeper at my training pictures, I'm figuring the problem is there being zero fucking consistency in the crown's appearance artist to artist, so the AI simply can't hack it even if it tries

Anonymous

Anonymous Wed 17 May 2023 03:27:47 No.7379620 Report

Quoted By:

resizing is cope for retards that can't train a decent model in the first place

Anonymous

Anonymous Wed 17 May 2023 03:29:29 No.7379623 Report

Quoted By: >>7379637

>>7379610
Once you learn how to use the jsons it's very edit to just make batch

Anonymous

Anonymous Wed 17 May 2023 03:30:01 No.7379624 Report

Quoted By: >>7379634

>>7379579
>https://kix.fanbox.cc/ (Source for image)
fuck off with your advertising

Anonymous

Anonymous Wed 17 May 2023 03:30:30 No.7379626 Report

Quoted By: >>7379628

>>7379613
No... it's not gonna require money to fix the model. It's how the images are GENERATED thats the problem. The model and training are fine.

Anonymous

Anonymous Wed 17 May 2023 03:31:43 No.7379628 Report

Quoted By:

>>7379626
>We finetune Stable Diffusion 2.1-base model on Laion dataset after changing the noise schedule to enforce zero terminal SNR and switching to the v prediction objec-tive.

Anonymous

Anonymous Wed 17 May 2023 03:32:42 No.7379629 Report

Quoted By: >>7379631

>>7379573
>We propose a few simple fixes: (1) rescale the noise schedule to enforce zero terminal SNR; (2) train the model with v prediction; (3)...
please learn to read

Anonymous

Anonymous Wed 17 May 2023 03:34:03 No.7379630 Report

Quoted By: >>7379636 >>7379905

>>7379615
1e4 unet with 5e5 TE and 128/128 on batch 1.
Though if you're struggling with some detail it might be better to try higher unet and lower dim/alpha. Havent tested enough to find one that doesnt smear eyes, but maybe 3e-4 to 7e-5 on 32/16.
Or put the most off model ones into a lower repeat folder

Anonymous

Anonymous Wed 17 May 2023 03:34:04 No.7379631 Report

Quoted By: >>7379633

>>7379629
imagine understanding any of that nerdspeak and not just endlessly proompting

Anonymous

Anonymous Wed 17 May 2023 03:35:27 No.7379633 Report

Quoted By:

>>7379631
meant for the other guy, whoops
you keep have fun, anon

Anonymous

Anonymous Wed 17 May 2023 03:36:03 No.7379634 Report

Quoted By: >>7379639 >>7379654

>>7379624
I am not the user you fucking idiot. Why should I post an image I didn't create without providing a source? This isn't deviantart.

Anonymous

Anonymous Wed 17 May 2023 03:36:28 No.7379636 Report

Quoted By: >>7379638 >>7379642 >>7379643 >>7379657 >>7379661

>>7379630
what are you finding for batch size? 4 is honestly coming in hot

Anonymous

Anonymous Wed 17 May 2023 03:36:41 No.7379637 Report

Quoted By:

>>7379623
Editting jsons is easy. Making the scripts run is the issue, I always get some error with no solutions on google.

Anonymous

Anonymous Wed 17 May 2023 03:37:10 No.7379638 Report

Quoted By:

>>7379636
If I were you I'd try 12.

Anonymous

Anonymous Wed 17 May 2023 03:37:23 No.7379639 Report

Quoted By:

>>7379634
>Why should I post an image I didn't create without providing a source?
lol got me

Anonymous

Anonymous Wed 17 May 2023 03:39:02 No.7379640 Report

Quoted By:

>>7379600
If you're actually an author of this paper I'm glad you're here with us anon

Anonymous

Anonymous Wed 17 May 2023 03:40:14 No.7379642 Report

Quoted By: >>7379643

>>7379636
The base model was trained on 2048 batches, so it must be the best.

Anonymous

Anonymous Wed 17 May 2023 03:41:13 No.7379643 Report

Quoted By:

>>7379642
true, more = better obviously. >>7379636 lower resolution if you cant train on 24

Anonymous

Anonymous Wed 17 May 2023 03:46:58 No.7379646 Report

Quoted By: >>7379650 >>7379770

>>7379608
quite a lot of people do. we don't really go by what people do and don't want to see when categorizing shit, otherwise a lot of the furry and dick girl shit would be sent to /b/ too lol. loli is a staple of hentai

Anonymous

Anonymous Wed 17 May 2023 03:52:04 No.7379650 Report

Quoted By: >>7379655 >>7379686

>>7379646
wow, good point. what would /h/ be without furries and futas

Anonymous

Anonymous Wed 17 May 2023 03:53:30 No.7379654 Report

Quoted By:

>>7379634
i will repeat. fuck off with your advertising

Anonymous

Anonymous Wed 17 May 2023 03:54:32 No.7379655 Report

Quoted By: >>7379770

>>7379650
nta but /h/'s definition of hentai is ridiculously narrow. You can't even post tentacles for fuck's sake. That's what, 40% of hentai?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00266-4095219020-best quality, ( (...).png, 473KiB, 768x768

Anonymous Wed 17 May 2023 03:55:03 No.7379656 Report

Quoted By: >>7379658

less talking more prompting

Anonymous

Anonymous Wed 17 May 2023 03:55:27 No.7379657 Report

Quoted By: >>7379661

>>7379636
you can't just adjust batch size without adjusting learning rates and expect good results

Anonymous

View Same Google ImgOps iqdb SauceNAO 00051-1289187061.png, 2MiB, 1024x1536

Anonymous Wed 17 May 2023 03:57:36 No.7379658 Report

Quoted By: >>7379660

>>7379656
fuck off my 1080's on its union mandated break

Anonymous

View Same Google ImgOps iqdb SauceNAO 00162-3373408734-riju, (dark ski (...).png, 1MiB, 768x1152

Anonymous Wed 17 May 2023 03:58:17 No.7379659 Report

Quoted By: >>7379661

>tfw there's not yet enough art of a new character to train a lora on

Anonymous

Anonymous Wed 17 May 2023 03:58:21 No.7379660 Report

Quoted By: >>7379662

>>7379658
>1080
lmao cheap ass rig you got there

Anonymous

Anonymous Wed 17 May 2023 03:59:29 No.7379661 Report

Quoted By: >>7379663 >>7379668 >>7379676 >>7379884

>>7379657
>>7379636
>struggles with having the lora learn a detail
>uses batch size greater than 1
Yep.

>>7379659
8 pics is enough

Anonymous

Anonymous Wed 17 May 2023 04:00:06 No.7379662 Report

Quoted By: >>7379666

>>7379660
jokes on you, I've got a 4090 it's just not plugged in

Anonymous

Anonymous Wed 17 May 2023 04:00:35 No.7379663 Report

Quoted By: >>7379664 >>7379671

>>7379661
anon i made the devolver-chan lora from 8 images, and i am telling you that you can't adjust batch size without changing learning rates.

Anonymous

Anonymous Wed 17 May 2023 04:02:23 No.7379664 Report

Quoted By: >>7379669

>>7379663
I never said you can.

Anonymous

Anonymous Wed 17 May 2023 04:02:25 No.7379665 Report

Quoted By: >>7380017

gayshit updated
>>7374813
>>7378824
Don't think it's ever happening, there are quite a few forks on gitgud, but they never got updated more than once.

Anonymous

Anonymous Wed 17 May 2023 04:02:53 No.7379666 Report

Quoted By: >>7379677

>>7379662
fuck! not busting my jaw anymore to get a better card again

Anonymous

View Same Google ImgOps iqdb SauceNAO 51282.jpg, 1MiB, 1680x1392

Anonymous Wed 17 May 2023 04:03:16 No.7379667 Report

Quoted By: >>7379735 >>7379809

prompt moms

Anonymous

View Same Google ImgOps iqdb SauceNAO 00157-3401751788-riju, dark skin (...).png, 1MiB, 768x1152

Anonymous Wed 17 May 2023 04:03:19 No.7379668 Report

Quoted By: >>7379677

>>7379661
I don't think there's even that many yet. And I'm not gonna be bothered to try and take screenshots off the switch.

Anonymous

Anonymous Wed 17 May 2023 04:03:36 No.7379669 Report

Quoted By: >>7379683

>>7379664
then i really don't get what your argument is, because i've worked with batch sizes up to 5 just fine

Anonymous

Anonymous Wed 17 May 2023 04:04:14 No.7379671 Report

Quoted By: >>7379673 >>7379681

>>7379663
People always say this but never mention any formula to scale LRs for batches.
Same for ranks and alphas.
I just found found two good settings for characters and styles respectively and don't touch them anymore.

Anonymous

Anonymous Wed 17 May 2023 04:05:18 No.7379673 Report

Quoted By: >>7379680

>>7379671
>People always say this but never mention any formula to scale LRs for batches.
because there is no formula. test and refine is all there is to it

Anonymous

View Same Google ImgOps iqdb SauceNAO 00027-2023044009.png, 2MiB, 1024x1536

Anonymous Wed 17 May 2023 04:06:03 No.7379675 Report

Quoted By: >>7380277 >>7388582

Anonymous

Anonymous Wed 17 May 2023 04:06:13 No.7379676 Report

Quoted By: >>7379683

>>7379661
1-3 seem underfit (1 often being severe gacha), 6+ seem overfit. the original paper literally recommends using multiples in batch sizes.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00149-3492718828.png, 2MiB, 1024x1536

Anonymous Wed 17 May 2023 04:06:18 No.7379677 Report

Quoted By:

>>7379668
it's a nintendo game lmao there'll be 400 pictures on twitter by the end of the week

>>7379666
time to collect tithes from the satanic temple

Anonymous

Anonymous Wed 17 May 2023 04:07:51 No.7379680 Report

Quoted By: >>7379685

>>7379673
Oh well, I don't have time to bake ten versions and run XYZs anymore.

Anonymous

Anonymous Wed 17 May 2023 04:08:10 No.7379681 Report

Quoted By:

>>7379671
>Same for ranks and alphas.
there is a formula for dim size/alpha. divide your alpha by your dimension. that's how much it scales your learning rate.

so 16/32 and 32/64 are the same, but 1/32 and 1/64 are different.

Anonymous

Anonymous Wed 17 May 2023 04:08:42 No.7379683 Report

Quoted By: >>7379685 >>7379694

>>7379669
Generalization.

>>7379676
what original paper? the one that recommends 2 - the lowest value they used?
>underfit
...? Just bake for more epochs? Adjust LRs? ???

Anonymous

Anonymous Wed 17 May 2023 04:10:55 No.7379685 Report

Quoted By: >>7379688

>>7379680
i mean if i had to do that i would just set up a script to bake them all in one shot at different learning rates then XYZ in the morning
>>7379683
>Generalization.
what, that loras above batch size 1 generally struggle to learn details? that's just false. with proper learning rates it works exactly the same as batch size 1

Anonymous

Anonymous Wed 17 May 2023 04:11:54 No.7379686 Report

Quoted By: >>7379689

>>7379650
valid nitpick, but the point i'm getting at is those aren't relegated strictly to /b/. loli belongs on h more so than on /b/ or /d/

Anonymous

Anonymous Wed 17 May 2023 04:12:40 No.7379688 Report

Quoted By: >>7379690

>>7379685
lol
not only a newfag but also a dimwit

Anonymous

Anonymous Wed 17 May 2023 04:13:24 No.7379689 Report

Quoted By: >>7380006

>>7379686
lolicon and or shota is literally rule one you can report things for on this board

Anonymous

Anonymous Wed 17 May 2023 04:13:50 No.7379690 Report

Quoted By: >>7379692

>>7379688
lol elaborate, because i've been here since before textual inversions were a thing buddy

Anonymous

Anonymous Wed 17 May 2023 04:14:24 No.7379691 Report

Quoted By: >>7379696 >>7379699 >>7379700 >>7379709 >>7379882

since when is recommending a batch size of 2-4 an unpopular opinion?

Anonymous

Anonymous Wed 17 May 2023 04:14:42 No.7379692 Report

Quoted By: >>7379695

>>7379690
then you should know about the guy who went above 20 bs.

Anonymous

Anonymous Wed 17 May 2023 04:15:01 No.7379694 Report

Quoted By:

>>7379683
>Generalization.
have you actually found out what this means yet

Anonymous

Anonymous Wed 17 May 2023 04:15:32 No.7379695 Report

Quoted By: >>7379697

>>7379692
>talks about up to batch size 5
>randomly brings up idiot that went to 20
lol. i know who i'm arguing with now

Anonymous

Anonymous Wed 17 May 2023 04:15:43 No.7379696 Report

Quoted By: >>7379698

>>7379691
>popularity
the popular opinion is to bake with 100 repeats. I dont think you should give a shit about whats popular

Anonymous

Anonymous Wed 17 May 2023 04:16:49 No.7379697 Report

Quoted By: >>7379704

>>7379695
>what, that loras above batch size 1 generally struggle to learn details? that's just false. with proper learning rates it works exactly the same as batch size 1
Its you who said they're exactly the same as batch size 1, are you backpedaling and saying they're not?

Anonymous

Anonymous Wed 17 May 2023 04:17:10 No.7379698 Report

Quoted By: >>7379702

>>7379696
we can all admit the damage that one youtuber has done is incalculable. If you have a high end card you can get 4x896, how is that supposed to be low quality?

Anonymous

Anonymous Wed 17 May 2023 04:18:00 No.7379699 Report

Quoted By:

>>7379691
it's one guy

Anonymous

Anonymous Wed 17 May 2023 04:18:52 No.7379700 Report

Quoted By:

>>7379691
same guy from a few threads ago shilling batch 1 because he's poor

Anonymous

Anonymous Wed 17 May 2023 04:19:57 No.7379702 Report

Quoted By: >>7379708

>>7379698
How is that supposed to not be low quality?

Anonymous

Anonymous Wed 17 May 2023 04:20:45 No.7379704 Report

Quoted By: >>7379716 >>7379741 >>7379746 >>7379756 >>7379822

>>7379697
i said that you can not make a generalization that higher batch size loras can't learn details, and specifically mentioned that i've tested up to batch size 5. with adjusted learning rates, they learn exactly the same way. i'm not going to say a ridiculous batch size like 20 would work fine, and you know that, but you're going to bring it up because you're clearly baiting. really miss when this thread wasn't full of idiots man

Anonymous

View Same Google ImgOps iqdb SauceNAO 00045-2053749658.png, 880KiB, 768x1152

Anonymous Wed 17 May 2023 04:22:59 No.7379707 Report

Quoted By: >>7382267

Anonymous

Anonymous Wed 17 May 2023 04:23:02 No.7379708 Report

Quoted By:

>>7379702
based retard

Anonymous

Anonymous Wed 17 May 2023 04:23:23 No.7379709 Report

Quoted By: >>7379711 >>7379714 >>7379715

>>7379691
since before newfags and schizos that never tried baking at 1 took over.
You can go back to like february when gradient got introduced for people with shitty GPUs to mimic batch sizes and the general opinion was not to bother with it because batch sizes sacrifice quality for speed. And gradient's slow as fuck.

Anonymous

Anonymous Wed 17 May 2023 04:23:53 No.7379711 Report

Quoted By:

>>7379709
>schizo hours

Anonymous

Anonymous Wed 17 May 2023 04:25:15 No.7379714 Report

Quoted By: >>7379717

>>7379709
> batch sizes sacrifice quality for speed
I'm sorry, but this is just cope from people who couldn't do actual higher batch sizes.

Anonymous

Anonymous Wed 17 May 2023 04:25:28 No.7379715 Report

Quoted By: >>7379717

>>7379709
>gradient
you're the only one talking about gradient checkpointing dipshit

Anonymous

Anonymous Wed 17 May 2023 04:26:29 No.7379716 Report

Quoted By: >>7379720 >>7379820

>>7379704
So if you can go up to batch size 5 and it's exactly as batch size 1 if you scale LRs, how come raising the batch size makes it suddenly stop with that behavior and change into something different? I.e if it doesn't change anything then raising the batch size infinitely would never degrade the lora as long as you had the correct LRs. Curious what kinda magic lies at the root of this transformation.

Anonymous

Anonymous Wed 17 May 2023 04:28:22 No.7379717 Report

Quoted By: >>7379723

>>7379715
am I not allowed to bring it up because it shatters your schizo theories?

>>7379714
are you retarded, this was said by people who actually used higher batch sizes at the time because poorfags were asking if its worth it.
then again theres even papers on it so obviously you are

Anonymous

Anonymous Wed 17 May 2023 04:29:12 No.7379720 Report

Quoted By: >>7379732

>>7379716
(you)

Anonymous

View Same Google ImgOps iqdb SauceNAO 00029-4181537556.png, 2MiB, 1024x1536

Anonymous Wed 17 May 2023 04:29:23 No.7379721 Report

Quoted By: >>7380206 >>7380277 >>7388582

Anonymous

Anonymous Wed 17 May 2023 04:29:56 No.7379723 Report

Quoted By: >>7379726 >>7379727 >>7379730 >>7379733

>>7379717
Oh. You're probably the same guy who has had the wrong opinion in these threads for months and you just shill papers.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00241-2422719167.png, 2MiB, 1024x1536

Anonymous Wed 17 May 2023 04:30:04 No.7379724 Report

Quoted By:

woah hey an image

Anonymous

Anonymous Wed 17 May 2023 04:31:12 No.7379726 Report

Quoted By:

>>7379723
True, papers bad. They'll never make me believe in the round earth bullshit

Anonymous

Anonymous Wed 17 May 2023 04:31:25 No.7379727 Report

Quoted By:

>>7379723
>you just shill papers
don't blame the papers, they don't even agree with his retarded conclusion

Anonymous

Anonymous Wed 17 May 2023 04:33:33 No.7379730 Report

Quoted By:

>>7379723
no that's me, but i shilled training at batch size 3 or 4 and only doing more than that if you could get a speed increase and really wanted it

Anonymous

Anonymous Wed 17 May 2023 04:34:06 No.7379732 Report

Quoted By: >>7379740

>>7379720
Yep, that's what I thought.

Anonymous

Anonymous Wed 17 May 2023 04:34:31 No.7379733 Report

Quoted By: >>7379739 >>7379746

>>7379723
>papers
not him, he couldn't produce even one when pressed about his retard fantasy. Just kept claiming he wouldn't spoonfeed

Anonymous

View Same Google ImgOps iqdb SauceNAO 00392-3604305491.png, 2MiB, 1024x1536

Anonymous Wed 17 May 2023 04:37:10 No.7379735 Report

Quoted By: >>7379809

>>7379667

Anonymous

View Same Google ImgOps iqdb SauceNAO 00029-2361697117.png, 2MiB, 1280x1280

Anonymous Wed 17 May 2023 04:39:09 No.7379738 Report

Quoted By:

Anonymous

Anonymous Wed 17 May 2023 04:39:18 No.7379739 Report

Quoted By: >>7379758

>>7379733
it was linked the last time this conversation came up retardbro

Anonymous

Anonymous Wed 17 May 2023 04:39:29 No.7379740 Report

Quoted By: >>7379744

>>7379732
i guess i really should be thanking you for outing yourself as bait earlier than expected. stopped me from wasting my energy

Anonymous

Anonymous Wed 17 May 2023 04:40:19 No.7379741 Report

Quoted By:

>>7379704
so if going up to 5 is equal to 1 why does it break at 10? genuinely curious

Anonymous

View Same Google ImgOps iqdb SauceNAO catbox_s5uufy.png, 944KiB, 768x1152

Anonymous Wed 17 May 2023 04:44:00 No.7379743 Report

Quoted By: >>7379777 >>7380277 >>7388582

Anonymous

Anonymous Wed 17 May 2023 04:45:44 No.7379744 Report

Quoted By:

>>7379740
>I have no argument so you're baiting
meds

Anonymous

Anonymous Wed 17 May 2023 04:47:53 No.7379746 Report

Quoted By: >>7379748 >>7379820

I too think batch size 5 is great. No I never trained on 1, why would I? I'm right because others agree with me.
Bet that's what 100 rep chads thought too.

>>7379733
Last time you retards talked about it the paper was linked, don't you have sdg to shitpost on?

>>7379704
Elaborate. Makes no sense that scaling would stop working at some point

Anonymous

Anonymous Wed 17 May 2023 04:48:49 No.7379748 Report

Quoted By: >>7379754

>>7379746
>new ip
lol

Anonymous

Anonymous Wed 17 May 2023 04:50:02 No.7379751 Report

Quoted By: >>7379752 >>7379753

Im new to lora training and i need help. I have questions on how to add activation keys for training.

What are steps needed in order to accomplish this. Is there a setting or extensions needed for this procedure.

Anonymous

Anonymous Wed 17 May 2023 04:51:13 No.7379752 Report

Quoted By: >>7379755

>>7379751
add a tag to what you're using to tag the images that stays at the front of the character (activation tags don't really work anyway)

Anonymous

Anonymous Wed 17 May 2023 04:52:31 No.7379753 Report

Quoted By: >>7379755

>>7379751
keeptokens=x
Where x is the number of tags you want kept.
Then just make sure those tags are at the front of your captions.
I'd stay away from this place atm; it's schizo hours

Anonymous

Anonymous Wed 17 May 2023 04:53:08 No.7379754 Report

Quoted By:

>>7379748
Not like he was banned if you're accusing of samefag.

Anonymous

Anonymous Wed 17 May 2023 04:54:32 No.7379755 Report

Quoted By: >>7379760

>>7379752
>>7379752
oh okay. I was wondering if it was something unique. In the wd1.4 tagger i would include additional tags and it would stay in the front of the txt file list, ao the only thing i missed was the keeptokens.

>>7379753
Thanks for the heads up and for the help. Would you reccomend using activation keys tags or just ignore this all together.

Anonymous

Anonymous Wed 17 May 2023 04:54:40 No.7379756 Report

Quoted By:

>>7379704
why would 20 not work fine if you adjust if generalization doesn't affect detail learning?

Anonymous

Anonymous Wed 17 May 2023 04:56:08 No.7379758 Report

Quoted By: >>7379759

>>7379739
bro I'm talking about the guy shilling batch 1, the small batch paper was posted in reply to me lol

Anonymous

Anonymous Wed 17 May 2023 04:59:38 No.7379759 Report

Quoted By: >>7379761

>>7379758
yeah as I was said - it was posted the last time this shit came up. what are you on about?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00002-2875761128.png, 2MiB, 1024x1536

Anonymous Wed 17 May 2023 04:59:38 No.7379760 Report

Quoted By:

>>7379755
>Would you reccomend using activation keys tags or just ignore this all together
unless you have a specific reason for wanting to use them then you don't really need them. also reminder to ignore the niggers spamming the thread and not posting images

Anonymous

Anonymous Wed 17 May 2023 05:00:39 No.7379761 Report

Quoted By: >>7379766

>>7379759
yeah it was posted, but not by batch 1 guy, which was my point

Anonymous

Anonymous Wed 17 May 2023 05:01:09 No.7379762 Report

Quoted By: >>7379768 >>7379774

>recommending no activation tag for chara LoRA
What's next, gonna tell him to prune tags like the guide we removed for being shit?

Anonymous

Anonymous Wed 17 May 2023 05:03:03 No.7379766 Report

Quoted By: >>7379771

>>7379761
And you know it wasn't him because..? Janny IP vision?
Also what the fuck are you going on about then, you said he's wrong because he didn't provide evidence, but you agree that evidence was posted, but since it was a different anon that guy is still wrong? What?

Anonymous

Anonymous Wed 17 May 2023 05:03:50 No.7379768 Report

Quoted By: >>7379772 >>7379790

>>7379762
wait we don't prune anymore either?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00287-195094962.png, 2MiB, 1024x1536

Anonymous Wed 17 May 2023 05:04:33 No.7379769 Report

Quoted By:

reading comprehension is hard

Anonymous

Anonymous Wed 17 May 2023 05:05:00 No.7379770 Report

Quoted By:

>>7379655
yeah, the lack of tentacles is rule I fucking question on /h/.
>>7379646
futas and furries aint on /h/ either. furries are relegated to /b/ and /trash/, thank god.

Anonymous

Anonymous Wed 17 May 2023 05:05:43 No.7379771 Report

Quoted By: >>7379779

>>7379766
>didn't provide evidence, but you agree that evidence was posted
there was no evidence posted. the papers this guy is claiming prove him right do not talk about batch size 1 or say that there's any benefit to it

Anonymous

Anonymous Wed 17 May 2023 05:07:31 No.7379772 Report

Quoted By: >>7380214

>>7379768
absolute state of this thread. Pruning wasnt a thing since lora training began outside of people bothered about using 2-3 tokens instead of 1. It hurts the lora's ability to change details

Anonymous

Anonymous Wed 17 May 2023 05:08:24 No.7379774 Report

Quoted By: >>7379775

>>7379762
If I don't want it activated I won't put it in the prompt you cocksucker

Anonymous

Anonymous Wed 17 May 2023 05:09:53 No.7379775 Report

Quoted By: >>7379776

>>7379774
he's saying to not have an activation tag trained in the first place, can you read?

Anonymous

Anonymous Wed 17 May 2023 05:11:02 No.7379776 Report

Quoted By:

>>7379775
no

Anonymous

View Same Google ImgOps iqdb SauceNAO 1665705266291.jpg, 41KiB, 405x720

Anonymous Wed 17 May 2023 05:11:44 No.7379777 Report

Quoted By: >>7379778

>>7379743
LORA drop?

Anonymous

Anonymous Wed 17 May 2023 05:12:31 No.7379778 Report

Quoted By: >>7379786

>>7379777
none of the loras listed in that catbox are new

Anonymous

Anonymous Wed 17 May 2023 05:13:13 No.7379779 Report

Quoted By: >>7379784

>>7379771
While its true we can't know for sure because it wasn't tried and I dont really agree with the papers as their criteria for what is best were probably different than ours, considering the lowest size they tested on was 2 and it "won", the conclusion is that less generalization = better and I dont see why 1 would be an exception to this.

Anonymous

Anonymous Wed 17 May 2023 05:15:28 No.7379782 Report

Quoted By: >>7379783 >>7379785

Aside from the guides at the top, what other LORA resources do you recommend that give great results?

Anonymous

Anonymous Wed 17 May 2023 05:16:25 No.7379783 Report

Quoted By:

>>7379782
https://rentry.org/59xed3

Anonymous

Anonymous Wed 17 May 2023 05:16:43 No.7379784 Report

Quoted By: >>7379788 >>7379798

>>7379779
lmao you seriously tried to claim evidence someone else posted as proving your point when you couldn't post your own?

Anonymous

Anonymous Wed 17 May 2023 05:16:50 No.7379785 Report

Quoted By: >>7379787

>>7379782
the spirits of my ancestors

Anonymous

View Same Google ImgOps iqdb SauceNAO 1676711179005.jpg, 29KiB, 353x358

Anonymous Wed 17 May 2023 05:18:01 No.7379786 Report

Quoted By:

>>7379778
Can they be found in the OP? Idk this computer shit im new to all this lmao, just started a couple hours ago

Anonymous

Anonymous Wed 17 May 2023 05:18:22 No.7379787 Report

Quoted By:

>>7379785
based

Anonymous

Anonymous Wed 17 May 2023 05:18:45 No.7379788 Report

Quoted By:

>>7379784
Oh so its just bait. I didnt know hdg started to shill bad settings on purpose, guess thats why everyone left or moved to the haven

Anonymous

Anonymous Wed 17 May 2023 05:20:01 No.7379790 Report

Quoted By: >>7379793 >>7379799 >>7379814 >>7379818

>>7379768
it can be useful in certain situations if you want to basically make the AI associate everything in an image with one token. i'll usually. if the dataset is large enough and examples exist then I usually take a handful of the best examples of something like an outfit and include a few pruned copies of it.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00007-3112991553.png, 2MiB, 1024x1536

Anonymous Wed 17 May 2023 05:20:52 No.7379793 Report

Quoted By: >>7379799

>>7379790
he wasn't actually asking a question anon

Anonymous

View Same Google ImgOps iqdb SauceNAO 00031-3502945310.png, 2MiB, 1024x1536

Anonymous Wed 17 May 2023 05:21:30 No.7379794 Report

Quoted By: >>7383861

Anonymous

View Same Google ImgOps iqdb SauceNAO 00077.jpg, 862KiB, 2000x4000

Anonymous Wed 17 May 2023 05:23:33 No.7379797 Report

Quoted By: >>7379810

Caulifla gets pretty wild

Anonymous

Anonymous Wed 17 May 2023 05:23:35 No.7379798 Report

Quoted By:

>>7379784
>going off on some other dude
Mindbroken. Take your medication

Anonymous

Anonymous Wed 17 May 2023 05:23:40 No.7379799 Report

Quoted By: >>7379804

>>7379790
Thanks.

>>7379793
I was asking a question you moron.

Anonymous

Anonymous Wed 17 May 2023 05:24:57 No.7379804 Report

Quoted By:

>>7379799
you honestly suck at this lol

Anonymous

View Same Google ImgOps iqdb SauceNAO 00028.jpg, 925KiB, 2000x4000

Anonymous Wed 17 May 2023 05:25:33 No.7379807 Report

Quoted By: >>7379810

Anonymous

Anonymous Wed 17 May 2023 05:25:48 No.7379809 Report

Quoted By:

>>7379667
>>7379735
catbox please?

Anonymous

Anonymous Wed 17 May 2023 05:26:00 No.7379810 Report

Quoted By: >>7379816

>>7379797
>>7379807
uh, your vae bro?

Anonymous

Anonymous Wed 17 May 2023 05:28:03 No.7379814 Report

Quoted By: >>7379818 >>7379826

>>7379790
Isn't that still worse than non pruned? I.e if you have "school uniform" and prune all tags related to it, it becomes harder to get a school uniform that's not the specific one it was trained on, meanwhile if you left tags you can put them in negs to prevent this and the "school uniform" token would still work if you set it not to shuffle/first

Anonymous

View Same Google ImgOps iqdb SauceNAO 00006.jpg, 868KiB, 2000x4000

Anonymous Wed 17 May 2023 05:29:28 No.7379816 Report

Quoted By:

>>7379810
yeh

Anonymous

Anonymous Wed 17 May 2023 05:29:45 No.7379818 Report

Quoted By: >>7379827

>>7379790
>>7379814
What about characters that have a unique costume? Wouldn't it be helpful to prune or at least standardize tags for the outfit?

Anonymous

Anonymous Wed 17 May 2023 05:31:34 No.7379820 Report

Quoted By: >>7379822

>>7379746
Last time I tried 2 it went worse than 1 since it overfit on random background stuff unprompted but not much difference otherwise. Overall felt like the issues people keep describing itt that I never ran into

>>7379716
Don't particularly wanna join the schizofest but what is this logic? It either always scales with no difference or there is difference and it's always present.

Anonymous

Anonymous Wed 17 May 2023 05:32:53 No.7379822 Report

Quoted By:

>>7379820
Meant for >>7379704

Anonymous

Anonymous Wed 17 May 2023 05:35:30 No.7379824 Report

Quoted By: >>7380732

>>7379553
more tiddy kanna please

Anonymous

View Same Google ImgOps iqdb SauceNAO 00034-174804298.png, 2MiB, 1408x1664

Anonymous Wed 17 May 2023 05:36:29 No.7379826 Report

Quoted By: >>7379829

fuck cats

>>7379814
do you not use kept tokens at all? the only use case for pruning things is to make your kept tokens stronger or trying to skew them in certain ways while retaining the whole dataset.

it usually isn't necessary but some of the tokens I use for outfits need that small amount of overfitting to reliably work

Anonymous

Anonymous Wed 17 May 2023 05:36:56 No.7379827 Report

Quoted By:

>>7379818
You always want to standardize. For these cases you just leave it tagged so the LoRA isn't overfit into generating only that piece of clothing when asking for one (like if you tag it as "skirt" generating any skirts that ain't that one would be challenging) OR prune it all and add a custom token but since adding custom tags people have to keep track of is a shit idea, I never tested this so can't say if there's issues

Anonymous

Anonymous Wed 17 May 2023 05:40:27 No.7379829 Report

Quoted By: >>7379847

>>7379826
I do keep them. I'm advocating for not pruning.
>outfits
Cant say I've had much of an issue with this, came down to settings and dataset usually. Though yes it will be a bit more reliable with pruning, but not worth it imo

Anonymous

Anonymous Wed 17 May 2023 05:42:15 No.7379831 Report

Quoted By: >>7381133

>>7379067
is there a lora for penis on face and head grab like this?

Anonymous

Anonymous Wed 17 May 2023 05:45:42 No.7379832 Report

Quoted By: >>7379835 >>7379837 >>7379849

Could some Anon help me out with this?
So my last a1111 pull was from last year and after updating I'm getting this following error at every turn

"A tensor with all NaNs was produced in VAE. This could be because there's not enough precision to represent the picture. Try adding --no-half-vae commandline argument to fix this. Use --disable-nan-check commandline argument to disable this check."

I'm on Paperspace so I'm not sure where I should put that commandline argument.

Anonymous

Anonymous Wed 17 May 2023 05:45:48 No.7379833 Report

Quoted By: >>7379836 >>7379839 >>7380163

>prompt: ass_focus, from_behind, huge_ass
>neg: face, cleavage, facing_viewer
>still getting front views
assbros...

Anonymous

View Same Google ImgOps iqdb SauceNAO 00048-449236549.png, 1MiB, 1264x1368

Anonymous Wed 17 May 2023 05:46:35 No.7379834 Report

Quoted By:

Anonymous

Anonymous Wed 17 May 2023 05:47:07 No.7379835 Report

Quoted By: >>7379839

>>7379832
Try adding --no-half-vae commandline argument to fix it. Use --disable-nan-check commandline argument to disable the check.

Anonymous

Anonymous Wed 17 May 2023 05:47:26 No.7379836 Report

Quoted By:

>>7379833
facing away
thank me later

Anonymous

Anonymous Wed 17 May 2023 05:49:52 No.7379837 Report

Quoted By: >>7379845 >>7379895 >>7380452

>>7379832
If it works similarly to the way the Colab notebook did, there should be a line that begins !COMMANDLINE_ARGS="
so put --no-half-vae inside the quote marks. But if you use a VAE other than the NAI VAE and its clones, you won't get this error.

Anonymous

Anonymous Wed 17 May 2023 05:51:12 No.7379839 Report

Quoted By:

>>7379833
disregard this retard >>7379835
pointless to use more VRAM when you could switch to a functional VAE like 840000 >>7379397

Anonymous

View Same Google ImgOps iqdb SauceNAO 00002-2279576463.png, 335KiB, 384x640

Anonymous Wed 17 May 2023 05:52:45 No.7379840 Report

Quoted By: >>7379841

this is apparently what my model thinks a capybara looks like.

Anonymous

Anonymous Wed 17 May 2023 05:53:39 No.7379841 Report

Quoted By:

>>7379840
I pull up

Anonymous

Anonymous Wed 17 May 2023 05:53:58 No.7379843 Report

Quoted By: >>7379902

>>7379075
shes so beautiful....

Anonymous

Anonymous Wed 17 May 2023 05:56:23 No.7379845 Report

Quoted By:

>>7379837
Thanks Anon, I'll try that. So much stuff to catch up on. I can only imagine how many different checkpoints have been created in this time

Anonymous

View Same Google ImgOps iqdb SauceNAO 00475-1801198088.png, 1MiB, 1024x1536

Anonymous Wed 17 May 2023 05:57:50 No.7379847 Report

Quoted By: >>7379855 >>7379878

>>7379829
Here's a practical example of why tag pruning is useful. If you load the Ghost in the Shell 1995 LoRA on gayshit and prompt "Kusanagi Motoko", it'll generate Kusanagi Motoko without you needing to describe her. From this, we can deduce that all of the tags describing Motoko's hairstyle and eye color were pruned when the LoRA was trained.

Anonymous

Anonymous Wed 17 May 2023 05:59:48 No.7379849 Report

Quoted By: >>7379853

>>7379832
"OutOfMemoryError: CUDA out of memory. Tried to allocate 16.00 GiB (GPU 0; 47.54 GiB total capacity; 18.37 GiB already allocated; 11.53 GiB free; 34.34 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. "

Nigga nani? This is what I get for pulling cold without checking.

Anonymous

Anonymous Wed 17 May 2023 06:01:15 No.7379853 Report

Quoted By: >>7379856

>>7379849
>48gb total capacity
nigger do you have two 4090s in there?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00217-129313473.png, 1MiB, 1024x1536

Anonymous Wed 17 May 2023 06:01:31 No.7379855 Report

Quoted By: >>7379859 >>7379878

>>7379847
But the Shirow Masamune LoRA on Civitai doesn't know who Motoko is, so you have to prompt "kusanagi motoko, purple hair, inverted bob". Even then, the "purple hair" tag will often influence the model to give her split bangs as if she were Katsuragi Misato from Evangelion. Don't try to add "ghost in the shell" as a prompt because you could end up generating an actual ghost, as shown here.

Anonymous

Anonymous Wed 17 May 2023 06:02:59 No.7379856 Report

Quoted By: >>7379858 >>7379860 >>7379864

>>7379853
No, just nvidia rtx a6000

Anonymous

Anonymous Wed 17 May 2023 06:05:05 No.7379858 Report

Quoted By: >>7379863

>>7379856
Right, you said you were on Paperspace. What resolution were you trying to generate at?

Anonymous

Anonymous Wed 17 May 2023 06:05:36 No.7379859 Report

Quoted By:

>>7379855
SPOOKED

Anonymous

View Same Google ImgOps iqdb SauceNAO 8b5829413ca6c323b77fafdeacafe48e.png, 121KiB, 984x700

Anonymous Wed 17 May 2023 06:06:27 No.7379860 Report

Quoted By: >>7379869

>>7379856
I hate richfags I HATE RICHFAGS

Anonymous

Anonymous Wed 17 May 2023 06:07:20 No.7379863 Report

Quoted By: >>7379869

>>7379858
Hires. fix 512x512 DPM++ 2M Karras, Upscaler Latent (nearest-exact). For whatever reason it can randomly create something and then spazzez out

Anonymous

Anonymous Wed 17 May 2023 06:07:34 No.7379864 Report

Quoted By:

>>7379856
Are quadros worth it or should I stick with geoforces

Anonymous

Anonymous Wed 17 May 2023 06:10:06 No.7379869 Report

Quoted By: >>7379873

>>7379860
He's not rich, he's renting it from Paperspace.

>>7379863
What resolution were you trying to upscale it to? I'm thinking that if it consumed so much memory, the resolution must have been really big.

Anonymous

Anonymous Wed 17 May 2023 06:12:04 No.7379873 Report

Quoted By: >>7379876

>>7379869
>What resolution were you trying to upscale it to? I'm thinking that if it consumed so much memory, the resolution must have been really big.

512x512 to 1024x1024 which is why I'm baffled

Anonymous

Anonymous Wed 17 May 2023 06:14:01 No.7379876 Report

Quoted By:

>>7379873
You could try adding --xformers to the command line arguments, and if it works correctly, it should reduce VRAM usage a lot.

Anonymous

Anonymous Wed 17 May 2023 06:16:25 No.7379878 Report

Quoted By: >>7379933

>>7379855
>>7379847
it's an incredible exception where one of the characters the model was actually trained on shares the full name and even then it's questionable - can you change Motoko's hairstyle and color in the gits lora? As for the artist one it might be just poorly trained/tagged, but it's not like not overfitting on Motoko is a bad thing - artist LoRAs tend to suffer from overfit on some chara.
>Gits in positives
No, you should be using the negative field.

Anonymous

Anonymous Wed 17 May 2023 06:20:19 No.7379882 Report

Quoted By:

>>7379691
Literally who the fuck cares about how other people are training their shit, just train your own shit according to your needs and if it works, use it.

Anonymous

Anonymous Wed 17 May 2023 06:21:46 No.7379884 Report

Quoted By: >>7379892 >>7379897

>>7379661
>batch size 1
why do I get the feeling that the people bitching about large batch sizes didn't bother to make sure that size was a divisor of their dataset
I used batch size 15, 768x768 res on a style lora and it came out great, despite me forgetting to change the settings from when I did BS 3

Anonymous

Anonymous Wed 17 May 2023 06:22:33 No.7379886 Report

Quoted By:

>runpod still doesn't accept non-custodial wallets for payments
fucking retards, who would want to use that garbage scam website to pay them

Anonymous

Anonymous Wed 17 May 2023 06:24:27 No.7379888 Report

Quoted By:

I want top post scat

Anonymous

Anonymous Wed 17 May 2023 06:24:47 No.7379889 Report

Quoted By: >>7379899

what are the best settings for training multi character lora?

Anonymous

Anonymous Wed 17 May 2023 06:26:09 No.7379891 Report

Quoted By:

>>7379430
take aspirin

Anonymous

Anonymous Wed 17 May 2023 06:26:50 No.7379892 Report

Quoted By: >>7379894

>>7379884
>didn't bother to make sure that size was a divisor of their dataset
>I used batch size 15
So, how many images were in your dataset?

Anonymous

Anonymous Wed 17 May 2023 06:27:23 No.7379894 Report

Quoted By:

>>7379892
75

Anonymous

Anonymous Wed 17 May 2023 06:27:51 No.7379895 Report

Quoted By: >>7379999

>>7379837
>But if you use a VAE other than the NAI VAE and its clones
Like for example?

Anonymous

View Same Google ImgOps iqdb SauceNAO tmpawmsna2a.png, 1MiB, 1024x1024

Anonymous Wed 17 May 2023 06:28:07 No.7379896 Report

Quoted By: >>7380051

Anonymous

Anonymous Wed 17 May 2023 06:28:09 No.7379897 Report

Quoted By: >>7379903

>>7379884
How would you get that idea? It's not like anyone complained about batch sizes overbaking, which they would due to ridiculous step amounts if this was the case. If anything batchfags would be more likely to try batch1 with insane LRs and unadjusted steps producing abominations, but it's not like they actually try it in the first place
>BS15
U sure u changed it to 15 so it wasn't still 3? Maybe you forgot.

Anonymous

Anonymous Wed 17 May 2023 06:29:10 No.7379899 Report

Quoted By: >>7379905

>>7379889
Same as usual?

Anonymous

View Same Google ImgOps iqdb SauceNAO 9996490989765690-891968717-killj (...).png, 2MiB, 1600x2200

Anonymous Wed 17 May 2023 06:29:37 No.7379902 Report

Quoted By: >>7379913

>>7379843
yea thats why i gen lots of killjoys

Anonymous

Anonymous Wed 17 May 2023 06:29:50 No.7379903 Report

Quoted By:

>>7379897
yep I'm sure
4090 is a monster and does unnatural things

Anonymous

Anonymous Wed 17 May 2023 06:30:00 No.7379905 Report

Quoted By: >>7379909

>>7379899

these ones are okay?
>>7379630

what about dadapt?

Anonymous

Anonymous Wed 17 May 2023 06:32:50 No.7379909 Report

Quoted By:

>>7379905
Should be? It's hard to fuck up 1e-4, try it and see if you run into issues?
>dadapt
Worse than Adam everytime I tried

Anonymous

Anonymous Wed 17 May 2023 06:36:05 No.7379911 Report

Quoted By: >>7379920 >>7379955

So is slower lower learning rate with more epochs better than fast learning rate with lower amount of epochs? I don't understand this shit man

Anonymous

Anonymous Wed 17 May 2023 06:38:33 No.7379913 Report

Quoted By: >>7379965 >>7379966

>>7379902
Put her in a wedding dress

Anonymous

Anonymous Wed 17 May 2023 06:39:05 No.7379914 Report

Quoted By:

>>7379235
The drawing itself for sure, coloring as well I'm not certain of, some people are very fast but that would be a minority.

Anonymous

Anonymous Wed 17 May 2023 06:45:39 No.7379920 Report

Quoted By:

>>7379911
lower with more epochs, since you can check epochs one by one for the sweet spot

Anonymous

View Same Google ImgOps iqdb SauceNAO 00048-3990503949.png, 1MiB, 1024x1536

Anonymous Wed 17 May 2023 06:55:04 No.7379933 Report

Quoted By: >>7379959

>>7379878
>can you change Motoko's hairstyle and color in the gits lora?
You can if you really crank up the emphasis:
<lora:gits1995:1> high quality, nsfw, kusanagi motoko, completely nude, (red hair:1.4), (parted bangs, long straight hair:1.9), cowboy shot, cyberpunk cityscape

Anonymous

Anonymous Wed 17 May 2023 06:58:16 No.7379939 Report

Quoted By: >>7380049

>>7379235
Yeah, you could use Photoshop to edit what you want and img to img it

Anonymous

View Same Google ImgOps iqdb SauceNAO xyz_grid-0002-3072572135.jpg, 536KiB, 1920x2419

Anonymous Wed 17 May 2023 06:58:40 No.7379940 Report

Quoted By:

>>7379354
Did a first draft a couple of days ago but don't have the will to finish it right now. Have to upscale some images, use the autotagger, and train it at 768 instead of 512.

Anonymous

Anonymous Wed 17 May 2023 06:59:17 No.7379942 Report

Quoted By:

What's the best paizuri lora? Or the best way to generate paizuri in general?

Anonymous

Anonymous Wed 17 May 2023 06:59:34 No.7379943 Report

Quoted By: >>7380010

Wow, after detailer does save a lot of time, only is 2-3s longer for regular prompting until you find what you're after.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00018-2285221146.png, 2MiB, 1024x1536

Anonymous Wed 17 May 2023 07:00:45 No.7379948 Report

Quoted By: >>7379961 >>7380068

loving new saya loras, enjoying 1e4-08 a lot but they're all great. i love saya

Anonymous

Anonymous Wed 17 May 2023 07:01:38 No.7379951 Report

Quoted By:

What's the consensus on deepfloyd? Couldn't it be great with something like Lora?

Anonymous

Anonymous Wed 17 May 2023 07:02:51 No.7379955 Report

Quoted By:

>>7379911
Neither afaik. Lower LR will not grasp details and overfit before it does. Higher LR will grasp them but have fun troubleshooting crisped eyes.

Anonymous

Anonymous Wed 17 May 2023 07:04:30 No.7379959 Report

Quoted By: >>7379970

>>7379933
>1.9
Holy kek, this here is why pruning's a thing. It's a miracle the entire image isn't fried, though I imagine it's had some negative effect regardless

Anonymous

Anonymous Wed 17 May 2023 07:05:51 No.7379961 Report

Quoted By: >>7379967 >>7380048

>>7379948
This post looks AI generated or like paid advertising. Are u ok? Nice image though

Anonymous

View Same Google ImgOps iqdb SauceNAO 9996490989765730-3899039903-kill (...).png, 3MiB, 1600x2400

Anonymous Wed 17 May 2023 07:12:44 No.7379965 Report

Quoted By: >>7379966 >>7379974

>>7379913
here you go. not sure if you want hat on or off so i did both

Anonymous

View Same Google ImgOps iqdb SauceNAO 9996490989765731-1409706302-kill (...).png, 3MiB, 1600x2400

Anonymous Wed 17 May 2023 07:13:20 No.7379966 Report

Quoted By:

>>7379913
>>7379965

Anonymous

Anonymous Wed 17 May 2023 07:13:24 No.7379967 Report

Quoted By: >>7379993 >>7380048

>>7379961
he shows up every now and then to derail the thread with his blog posts. Just ignore like everyone else does

Anonymous

Anonymous Wed 17 May 2023 07:13:37 No.7379968 Report

Quoted By: >>7379973 >>7379975

new fag here please share some advice

3060 12gb

or

3060 ti 8gb ?

for heavy big prompts and many loras + hires fix

thanks in advance

Anonymous

View Same Google ImgOps iqdb SauceNAO 00077-3978263032.png, 1MiB, 1024x1536

Anonymous Wed 17 May 2023 07:14:57 No.7379970 Report

Quoted By: >>7379994

>>7379959
That's just an experiment, though. If I really wanted to generate a character who doesn't have Motoko's hairstyle, I wouldn't put Motoko's name in the prompt at all:
<lora:gits1995:1> high quality, 1girl, completely nude, parted bangs, long straight red hair, cowboy shot, cyberpunk cityscape

Anonymous

Anonymous Wed 17 May 2023 07:15:07 No.7379973 Report

Quoted By: >>7379976

>>7379968
12 is bigger than 8

Anonymous

View Same Google ImgOps iqdb SauceNAO 00031-645912231.png, 1MiB, 1024x1280

Anonymous Wed 17 May 2023 07:15:15 No.7379974 Report

Quoted By: >>7379985

>>7379965
i capture but a fraction of your artistry...

Anonymous

Anonymous Wed 17 May 2023 07:15:33 No.7379975 Report

Quoted By:

>>7379968
12gb

Anonymous

Anonymous Wed 17 May 2023 07:15:58 No.7379976 Report

Quoted By: >>7379978 >>7379980 >>7379981 >>7379986 >>7379996

>>7379973
but isn't the 3060 ti faster ?

Anonymous

Anonymous Wed 17 May 2023 07:17:03 No.7379978 Report

Quoted By:

>>7379976
doesn't matter how fast it is if you OOM

Anonymous

Anonymous Wed 17 May 2023 07:18:01 No.7379980 Report

Quoted By:

>>7379976
The speed difference is negligible, but the additional VRAM makes a lot of difference, especially if you'll ever want to start making your own loras

Anonymous

Anonymous Wed 17 May 2023 07:18:04 No.7379981 Report

Quoted By:

>>7379976
YOU CAN'T COOM IF YOU OOM

Anonymous

View Same Google ImgOps iqdb SauceNAO 1680927978703261-3353086314.png, 2MiB, 1024x1456

Anonymous Wed 17 May 2023 07:19:59 No.7379983 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO 9996490989765654-1710591105-kill (...).png, 1MiB, 1280x1600

Anonymous Wed 17 May 2023 07:20:45 No.7379985 Report

Quoted By:

>>7379974
ayo? thats pretty sick. now we can enjoy even more killjoys. just be aware that im going to ask catboxes here and there for killjoys

Anonymous

Anonymous Wed 17 May 2023 07:23:03 No.7379986 Report

Quoted By: >>7379989

>>7379976
If you want to make loras and several images at the same time go with the 3060. If you want to make few images fast and have a overall better card for vidya go for the ti.

Anonymous

Anonymous Wed 17 May 2023 07:24:21 No.7379989 Report

Quoted By: >>7379995

>>7379986
>vidya
what the fuck is that? You can use your graphics card for shit other than prompting/training?

Anonymous

View Same Google ImgOps iqdb SauceNAO 111.png, 1MiB, 1024x1024

Anonymous Wed 17 May 2023 07:25:09 No.7379992 Report

Quoted By: >>7380017

gayshit anon is alive...

Anonymous

Anonymous Wed 17 May 2023 07:26:58 No.7379993 Report

Quoted By: >>7380012 >>7380033

>>7379967
You're derailing more than he's ever did, schizoid

Anonymous

Anonymous Wed 17 May 2023 07:28:07 No.7379994 Report

Quoted By:

>>7379970
That tends to result in details being fucked though since it's trained on the names - like her eyes are a bit off here

Anonymous

Anonymous Wed 17 May 2023 07:28:34 No.7379995 Report

Quoted By:

>>7379989
I personally use my 2k gaming computer for prompting, garbage h-games, non resource intensive games like eu4 and homm3, and dota. No, I'm not a slav

Anonymous

Anonymous Wed 17 May 2023 07:30:08 No.7379996 Report

Quoted By:

>>7379976
12 GB is nice but retards undermine the speed. Look up graphs on /g/. Sure you won't oom at 1400x1400 at 12GB but who the fuck wants to wait ages for the image - if speed isn't a concern you can --medvram or --lowvram 8gb to still get that res.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00103-679626472.png, 1MiB, 1024x1536

Anonymous Wed 17 May 2023 07:33:18 No.7379999 Report

Quoted By: >>7380014

>>7379895
Like 84000, which I mentioned in >>7379397.

Anonymous

View Same Google ImgOps iqdb SauceNAO sdsdfe.png, 556KiB, 512x824

Anonymous Wed 17 May 2023 07:38:26 No.7380002 Report

Quoted By:

I gave in and trying to wrangle out a specific style out of a model, and went with the advertised one.
Things get so much easier to prompt.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00027.png, 2MiB, 1024x1648

Anonymous Wed 17 May 2023 07:39:54 No.7380004 Report

Quoted By:

I gave in and trying to wrangle out a specific style out of a model, and went with the advertised one.
Things get so much easier to prompt.

Anonymous

Anonymous Wed 17 May 2023 07:40:16 No.7380006 Report

Quoted By:

>>7379689
yeah cool, but we're discussing why that's a rule in the first place

Anonymous

Anonymous Wed 17 May 2023 07:41:08 No.7380007 Report

Quoted By:

So... What models are you guys using?

Anonymous

Anonymous Wed 17 May 2023 07:43:25 No.7380010 Report

Quoted By: >>7380015

>>7379943
>after detailer
What's that?

Anonymous

Anonymous Wed 17 May 2023 07:45:57 No.7380012 Report

Quoted By:

>>7379993
sorry, still not paying for early access

Anonymous

Anonymous Wed 17 May 2023 07:46:46 No.7380014 Report

Quoted By:

>>7379999
Thanks, I'll download it.

Anonymous

Anonymous Wed 17 May 2023 07:47:10 No.7380015 Report

Quoted By:

>>7380010
Extension that basically scans the image for a subject, inpaints it and applies a hires fix(I believe).
So when you're prompting 1girl, like 99% of us in here are, it can speed up the process.

Anonymous

View Same Google ImgOps iqdb SauceNAO 1681709626768384.png, 636KiB, 512x1024

Anonymous Wed 17 May 2023 07:47:26 No.7380016 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO 00006-394593430.png, 2MiB, 1024x1280

Anonymous Wed 17 May 2023 07:48:30 No.7380017 Report

Quoted By:

>>7379665
>>7379992
Plenty of stuff still didn't get uploaded there sadly

Anonymous

Anonymous Wed 17 May 2023 07:52:22 No.7380019 Report

Quoted By: >>7380020 >>7380029

>got the perfect picture
>it's 512x512
How do I upscale it properly while adding more details?

Anonymous

Anonymous Wed 17 May 2023 07:53:39 No.7380020 Report

Quoted By:

>>7380019
Use the same seed, apply hires fix, denoise below 0.5, fiddle around, hope for the best.

Anonymous

Anonymous Wed 17 May 2023 07:55:54 No.7380023 Report

Quoted By: >>7380049

>>7379229
>>7379235
I'm always impressed by how persistent you seem to be when it comes to inpainting, I rarely bother going beyond 5-10 attempts. I can respect spending lots of time on one pic instead of continuously printing out new ones though.
Do you not use any sort of image editing whatsoever? Not saying it'd be a complete replacement to just rerolling in img2img/inpainting but I personally find it pretty helpful to nudge the AI towards what I have in mind.

Anonymous

Anonymous Wed 17 May 2023 07:57:08 No.7380025 Report

Quoted By: >>7380027

civtai blocking my image because it has "loli" in the generation prompt.

it's over...

Anonymous

Anonymous Wed 17 May 2023 08:00:10 No.7380027 Report

Quoted By:

>>7380025
it's what you get for cloutposting

Anonymous

View Same Google ImgOps iqdb SauceNAO 00015-3186994823.png, 1MiB, 1024x1536

Anonymous Wed 17 May 2023 08:01:29 No.7380029 Report

Quoted By: >>7380154

>>7380019
Add "upscaler_for_img2img" to your quicksettings
Check "With img2img, do exactly the amount of steps the slider specifies (normally you'd do less with less denoising)." setting

Use img2img, type in the desired resolution manually (so if it's 2x it would be 1024x1024 for your case)
Choose the img2img upscaler, I suggest 4x-animesharp or remacri
Use 0.5-0.55 denoise if you want it to be detailed and ready to redo the upscale several times in case it spergs out, 0.35-0.45 if you want it to look more like initial gen but a bit undetailed

That's my usual upscale process on a9fed7c3. Reusing hires fix with the same seed is also an option but I feel like it gives you less control over upscaling and it kinda wastes time if you are a GPUlet

Anonymous

Anonymous Wed 17 May 2023 08:05:13 No.7380033 Report

Quoted By: >>7380061 >>7380066

>>7379993
>he
think you mean you, lmao

Anonymous

View Same Google ImgOps iqdb SauceNAO yoshimon-1061it-novae.png, 2MiB, 1024x1536

Anonymous Wed 17 May 2023 08:07:53 No.7380040 Report

Quoted By: >>7380041

https://mega.nz/folder/ctl1FYoK#BYlpywutnH4psbEcxsbLLg
added yoshimon
>>7379470
>>7379479
here's the lora

Anonymous

View Same Google ImgOps iqdb SauceNAO 00093-2614442089.png, 1MiB, 1024x1536

Anonymous Wed 17 May 2023 08:08:35 No.7380041 Report

Quoted By:

>>7380040

Anonymous

View Same Google ImgOps iqdb SauceNAO 00272-1259007451.png, 2MiB, 1024x1536

Anonymous Wed 17 May 2023 08:09:37 No.7380044 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO 00073-1041443006.png, 1MiB, 1064x1064

Anonymous Wed 17 May 2023 08:14:10 No.7380048 Report

Quoted By: >>7380068

>>7379961
thanks. and i'm just happy being able to gen things i love and i'm just expressing my appreciation

>>7379967
(You) dropped something

Anonymous

Anonymous Wed 17 May 2023 08:14:48 No.7380049 Report

Quoted By:

>>7379939
>>7380023
Tthis isn't purely rolling though it might seem that way from the way the image sequence plays out.
I do a fair bit of editing, depends on how obscure the goal is.
Here's the process doc, it'll be missing a few frames since I tend to edit destructively.
https://www.mediafire.com/file/1nalo2z58glcffc/Synchronicity.psd/file

I converted from Sai2 to PSD so it may not be 1:1.

Anonymous

Anonymous Wed 17 May 2023 08:17:54 No.7380051 Report

Quoted By:

>>7379896
cat box plis?

Anonymous

View Same Google ImgOps iqdb SauceNAO catbox_olfe8d.png, 1MiB, 1024x1024

Anonymous Wed 17 May 2023 08:21:32 No.7380053 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO 1675518577254285.jpg, 331KiB, 1408x1792

Anonymous Wed 17 May 2023 08:22:08 No.7380055 Report

Quoted By: >>7380056 >>7380058

Help I've fallen into the (traditional media:1.3) hole and I can't get out.

Anonymous

Anonymous Wed 17 May 2023 08:22:52 No.7380056 Report

Quoted By:

>>7380055
moarrrr

Anonymous

View Same Google ImgOps iqdb SauceNAO 00269-4078593046.png, 2MiB, 1024x1536

Anonymous Wed 17 May 2023 08:24:23 No.7380058 Report

Quoted By:

>>7380055
catbox? nice gen

Anonymous

Anonymous Wed 17 May 2023 08:26:36 No.7380061 Report

Quoted By: >>7380064

>>7380033
>everyone is the same person
Sure. Do your stalking reps cause you're lacking

Anonymous

Anonymous Wed 17 May 2023 08:27:36 No.7380064 Report

Quoted By:

>>7380061
>new ip again
lol

Anonymous

Anonymous Wed 17 May 2023 08:28:04 No.7380066 Report

Quoted By:

>>7380033
oh no bro he's gonna call you a schizo

Anonymous

Anonymous Wed 17 May 2023 08:29:09 No.7380068 Report

Quoted By: >>7380217

>>7379948
>>7380048
Good to know it's working. Do tell if there's some issues, like being worse compared to the previous or so

Anonymous

View Same Google ImgOps iqdb SauceNAO 00189-1380383246-killjoy, round (...).png, 1MiB, 768x1152

Anonymous Wed 17 May 2023 08:29:15 No.7380069 Report

Quoted By:

Why does the AI spaz out about naked aprons.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00373-3541736782-nsfw, scan, (tr (...).png, 2MiB, 1024x1536

Anonymous Wed 17 May 2023 08:33:22 No.7380074 Report

Quoted By: >>7380089 >>7380128

>>7379422
witchbox: https://files.catbox.moe/eimrdn.png

Anonymous

View Same Google ImgOps iqdb SauceNAO 00020-1675965131.png, 2MiB, 1024x1536

Anonymous Wed 17 May 2023 08:42:45 No.7380089 Report

Quoted By:

>>7380074
ty

Anonymous

Anonymous Wed 17 May 2023 08:55:07 No.7380109 Report

Quoted By: >>7380115 >>7380119

>samefag avatarfag with his acog tier boring images
>shitposters trying to gaslight
>almost no one makes LoRAs anymore
Sad. Maybe it's time to leave, even the OP's a joke

Anonymous

Anonymous Wed 17 May 2023 08:59:58 No.7380115 Report

Quoted By:

>>7380109
I agree with you but yoshimon just dropped

Anonymous

Anonymous Wed 17 May 2023 09:02:49 No.7380119 Report

Quoted By: >>7380126

>>7380109
Things has gotten pretty stale, the tech isn't really improving, there's not much enthusiasm anymore. Maybe we can pray that there's another major leak happening soon

Anonymous

Anonymous Wed 17 May 2023 09:05:31 No.7380123 Report

Quoted By: >>7380637 >>7380686

What consumer gpu would be similar in generation speed to colab free tier, if any?

Anonymous

Anonymous Wed 17 May 2023 09:09:08 No.7380126 Report

Quoted By: >>7380127 >>7380130 >>7380331

>>7380119
Being able to make mostly anything you want turned out more boring than expected. As with all things

Anonymous

Anonymous Wed 17 May 2023 09:11:19 No.7380127 Report

Quoted By: >>7380133 >>7380331

>>7380126
Yeah, I spend hours crafting a lora only to use it twice before I get bored.

Anonymous

Anonymous Wed 17 May 2023 09:11:51 No.7380128 Report

Quoted By: >>7380129

>>7380074
catbox for this one?

Anonymous

Anonymous Wed 17 May 2023 09:14:12 No.7380129 Report

Quoted By:

>>7380128
https://files.catbox.moe/79iaec.png

Anonymous

Anonymous Wed 17 May 2023 09:15:00 No.7380130 Report

Quoted By: >>7380141

>>7380126
It's still not really that boring to me, but I joined pretty late.
>Being able to make mostly anything you want turned out more boring than expected
SD is far from perfect tbqh. But yeah you can do a lot of stuff with enough persistence, but it doesn't pay off most of the time because most normaloids just seethe at the results no matter what and call you a thief even if you don't intend to monetize stuff. And simply getting (You)s at hdg might have become not enough for some of the effort posters

Anonymous

Anonymous Wed 17 May 2023 09:19:17 No.7380133 Report

Quoted By:

>>7380127
Same here mostly although it's usually because just as I finish up a lora I come up with an idea for the next one

Anonymous

Anonymous Wed 17 May 2023 09:22:23 No.7380136 Report

Quoted By: >>7380142

Anonymous

Anonymous Wed 17 May 2023 09:26:19 No.7380141 Report

Quoted By: >>7380150

>>7380130
AI clout is worthless since there isn't much you can talk about or impress people with - it's not organic, only people that could turn this into a valuable interaction would be guys like inpaint chad - sharing his workflow webms if Twitter wasnt anal towards nsfw

Anonymous

Anonymous Wed 17 May 2023 09:27:20 No.7380142 Report

Quoted By: >>7380252

>>7380136
Was so much better, just needed to be trimmed instead of fucking vandalized

Anonymous

Anonymous Wed 17 May 2023 09:31:37 No.7380150 Report

Quoted By: >>7380162

>>7380141
>it's not organic
Garbage inkcel luddite argument. If the end-result is actually good, it shouldn't matter how it was made. 90% of AI art being garbage due to NAI grifters is a whole other issue.

Anonymous

Anonymous Wed 17 May 2023 09:34:07 No.7380154 Report

Quoted By:

>>7380029
NTA, but that quicksetting helps a ton, thanks.

Anonymous

Anonymous Wed 17 May 2023 09:38:52 No.7380162 Report

Quoted By: >>7380182

>>7380150
I mean yeah it looks good and stuff, but im talking about interaction.
There's no reason for people to really interact with (You), most you'll get is people asking for prompts if metadata's stripped.
It also doesn't hit the same way, lewds drawn by humans are still superior if done well and AI in general is less impressive than a masterpiece done by hand. Keep in mind I mean it for interaction and social aspects only, stuff like people commenting on it more often etc.
Im sure you've seen people talk about AI accounts getting less interactions here over time, where they'd just get likes or reposts and thats about it

Anonymous

Anonymous Wed 17 May 2023 09:39:37 No.7380163 Report

Quoted By:

>>7379833
First stop using underscores, they don't actually matter and they waste tokens. Second, just put from behind at a higher weight. Usually 1.2 or 1.4 does the job.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00442-4076774475.jpg, 757KiB, 2048x3072

Anonymous Wed 17 May 2023 09:42:29 No.7380171 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO 8279dcef8832c4844ef21d191dbc5fdf.png, 600KiB, 679x335

Anonymous Wed 17 May 2023 09:44:24 No.7380174 Report

Quoted By:

>left the denoise at too high

Anonymous

Anonymous Wed 17 May 2023 09:48:58 No.7380182 Report

Quoted By: >>7380196

>>7380162
I've seen some AI pixiv accounts that get plenty of interaction tbqh.
>There's no reason for people to really interact with (You), most you'll get is people asking for prompts if metadata's stripped.
And what's the reason to interact with an artist..? Despite AI being obviously magnitudes easier to do compared to traditional/digital art, people are still unwilling to mess around with it at all. Most users are just extremely tech illiterate and even installing git would be a feat for them, believe me.
>It also doesn't hit the same way, lewds drawn by humans are still superior if done well and AI in general is less impressive than a masterpiece done by hand.
Of course, top notch artists are still going to be better than some amateur cumbrain typing prompts and just posting result without further retouch. But the fact that they can even compare is still very impressive. Even basic understanding of anatomy/perspective and other are useful to make your outputs stand out.
>Im sure you've seen people talk about AI accounts getting less interactions here over time, where they'd just get likes or reposts and thats about it
That'd be the case for 95% of digital artists I imagine. You simply never heard of them.

Anonymous

Anonymous Wed 17 May 2023 09:49:15 No.7380183 Report

Quoted By:

>>7379397
>Based65
thanks, that fixed all my problems

Anonymous

Anonymous Wed 17 May 2023 10:02:02 No.7380196 Report

Quoted By:

>>7380182
>That'd be the case for 95% of digital artists I imagine. You simply never heard of them.
people meant popular AI accounts having no interaction.
>reason to interact with an artist
commenting on some details (while I know you can get detailed with AI, alot of it is often left unplanned beyond some general outline) or stylistic choice - for example using a lora of x chara to draw the chara in OG style is different from bending your own style to make it look more like the OG.
>top notch artists
Its not about the image quality, its the process, the mere fact of it being hand drawn is more impressive in the human sense. Similarly to someone using a chainsaw to cut down a forest doesn't elicit the same emotional response as if someone manually did it with an axe. I may like AI more, but im more likely to think "wow" and leave some token of appreciation at a manually drawn one.

Anonymous

Anonymous Wed 17 May 2023 10:09:27 No.7380200 Report

Quoted By: >>7380208 >>7380223

Anyone know why images sometimes, and even often depending on the prompt, produce a layer of mist at the bottom edge of the image? Is it just overfitting? And would there be a way to solve this that doesn't involve putting fog, mist in the negatives?

Anonymous

Anonymous Wed 17 May 2023 10:15:01 No.7380204 Report

Quoted By:

>>7378888
this style https://pixeldrain com/u/LBhcw8E5, thank you brotherman

Anonymous

View Same Google ImgOps iqdb SauceNAO 1674530874384.jpg, 69KiB, 702x1024

Anonymous Wed 17 May 2023 10:17:41 No.7380206 Report

Quoted By: >>7380648

>>7379721
VAE, My dude?

Anonymous

Anonymous Wed 17 May 2023 10:19:18 No.7380208 Report

Quoted By: >>7380215

>>7380200
its the lewd mist from "nsfw" prompt usually

Anonymous

Anonymous Wed 17 May 2023 10:27:41 No.7380214 Report

Quoted By:

>>7379772
Funny how one retard can fuck up communities for months, eh?/Same with the youtube guide guy.

Anonymous

Anonymous Wed 17 May 2023 10:28:03 No.7380215 Report

Quoted By: >>7380221

>>7380208
I have actually never used that token before. I do often prompt for scenery though so that would be my guess (overfit towards producing mist on the edge of images because artists like doing that).

Anonymous

View Same Google ImgOps iqdb SauceNAO 00117-103961569.png, 2MiB, 1264x1264

Anonymous Wed 17 May 2023 10:28:34 No.7380217 Report

Quoted By: >>7380244

>>7380068
the new ones make things look more normal in most of my abstract gens and they seem to have lost a bit of meatiness alongside the 3d cg look but i don't consider that to be an issue, old ones are perfect for that already. i really like how the new ones especially the 08 one are so responsive to clothing prompts and they don't seem to force her og outfit with the ribbon or something random on her chest as much, it happens quite rarely now and it's an easy fix anyway. they work much better with other character loras as well. so definitely not worse, just more saya focused, cleaner and easier to use

you've made a lot of grids yourself so this is probably not needed or useful but i'll just leave it here, i was trying to see for myself how all of them look alongside each other. idk what saya1 is though lol
https://files.catbox.moe/n07r9e.png
https://files.catbox.moe/hl30uk.png
https://files.catbox.moe/9uccry.png
https://files.catbox.moe/prhsq4.png
https://files.catbox.moe/up5cut.png

also saya style hanae is cute
https://files.catbox.moe/dwzg0r.png
https://files.catbox.moe/tz95br.png
https://files.catbox.moe/y413ro.png

Anonymous

Anonymous Wed 17 May 2023 10:31:00 No.7380221 Report

Quoted By: >>7380269

>>7380215
Its one of the thing AOM poisoned the well with. You can try putting "nsfw" in negs maybe if nothing else helps, unsure. Maybe 'steaming body"

Anonymous

View Same Google ImgOps iqdb SauceNAO s_xyz_grid-0934-2377997607-upper (...).jpg, 1MiB, 2880x5052

Anonymous Wed 17 May 2023 10:31:25 No.7380222 Report

Quoted By: >>7380231

i did a bunch of tests and they were all like this, i'm feeling that bucket steps are another lever you shouldn't pull like no upscale, at least for small features like this. still more tests to go

Anonymous

Anonymous Wed 17 May 2023 10:32:15 No.7380223 Report

Quoted By: >>7380269

>>7380200
Try "depth of field, blurry, steam, bokeh" in negs

Anonymous

Anonymous Wed 17 May 2023 10:39:37 No.7380230 Report

Quoted By:

>>7379055
i'll make one after i'm satisfied testing but they keep adding new fucking training features

Anonymous

Anonymous Wed 17 May 2023 10:40:12 No.7380231 Report

Quoted By: >>7380242 >>7380418

>>7380222
are these different training seeds?

Anonymous

Anonymous Wed 17 May 2023 10:48:00 No.7380242 Report

Quoted By: >>7380246 >>7380249 >>7380418

>>7380231
no and why do you ask that question, what kind of idiot would test training while randomizing their training

Anonymous

View Same Google ImgOps iqdb SauceNAO maiden rape assault violent seme (...).png, 9KiB, 83x94

Anonymous Wed 17 May 2023 10:50:12 No.7380244 Report

Quoted By: >>7380289

>>7380217
>grids
Helpful. The new ones came out surprisingly weak. I'd consider a rebake but at this point theres enough versions that its probably redundant. Are these all img2img btw?
>3D CG
I wonder if "game CG" would get that on track again, considering its a decently strong tag.
>meatiness
did you try messing with weights? I noticed it wasn't as meaty but didn't try raising horror etc to see if it changes anything
>Hanae
Given how many flowers that lora can generate I wonder if you can mix that for some neat meat flora.
>pic
do you have the OG? Seems familiar. I assume its not tx2img.

Anonymous

Anonymous Wed 17 May 2023 10:54:56 No.7380246 Report

Quoted By: >>7380249

>>7380242
The average hdg user probably.

Anonymous

Anonymous Wed 17 May 2023 10:55:53 No.7380249 Report

Quoted By: >>7380260

>>7380242
>>7380246
this desu

Anonymous

Anonymous Wed 17 May 2023 10:58:00 No.7380252 Report

Quoted By:

>>7380142
DIY

Anonymous

Anonymous Wed 17 May 2023 11:07:50 No.7380260 Report

Quoted By:

>>7380249
https://www.youtube.com/watch?v=bt4e_ZcO1vY

Anonymous

View Same Google ImgOps iqdb SauceNAO 00473-nsfw, scan.png, 2MiB, 1024x1536

Anonymous Wed 17 May 2023 11:14:32 No.7380264 Report

Quoted By: >>7380266

Anonymous

Anonymous Wed 17 May 2023 11:15:56 No.7380266 Report

Quoted By: >>7380270

>>7380264
cute butt

Anonymous

Anonymous Wed 17 May 2023 11:19:07 No.7380269 Report

Quoted By:

>>7380223
>>7380221
Trying it out and it still seems to happen, but maybe not as often? Would need to do more testing.

Anonymous

Anonymous Wed 17 May 2023 11:19:37 No.7380270 Report

Quoted By:

>>7380266
Yes and a cute art style.

Anonymous

View Same Google ImgOps iqdb SauceNAO d6d18c18eb3bdf8f0501d8e2b736ebaf.png, 2MiB, 2048x1536

Anonymous Wed 17 May 2023 11:23:05 No.7380276 Report

Quoted By:

Found this on rule34.xxx it seems like someone managed to make a Sucy Manbavaran LoRA

Anonymous

View Same Google ImgOps iqdb SauceNAO Capture.png, 219KiB, 838x266

Anonymous Wed 17 May 2023 11:23:06 No.7380277 Report

Quoted By: >>7380648

>>7379721
>>7379675
>>7379593
>>7379743
Can I use this girl for an F-List profile? Do you have any blowjob art generated? I've been saving images for a bit but I don't know if I have very many oral focused images. She's really pretty and I'd like to make a profile around her.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00047-3844815433.png, 2MiB, 1240x1240

Anonymous Wed 17 May 2023 11:36:12 No.7380289 Report

Quoted By: >>7380422

>>7380244
>that filename
fucking kek, doushin fd would be amazing

grids are just txt2img prompts, they should all have metadata, and i guess new loras are weaker than the older ones but they do have the style and most importantly they get saya right, at least to me. hair flaps are not present in some because i didn't prompt them, a few of those grids were just old gens so i didn't change the prompt

>meat cg
i haven't tried "game CG" specifically, all i did was throw in my older 3d-ish gens and compared them one by one. but you're right, i'm gonna try that tag, mess with the weights and see what happens.
meat flower hanae sounds exciting. another one to the list of to-do gens, thanks

>og pic
yeah that one that was img2img, here's the og
https://e.snmc.io/i/1200/s/a2be58a1b744d0a10b640fb9f68449c0/5335230
picrel is img2img too, i like forcing colors that way

Anonymous

View Same Google ImgOps iqdb SauceNAO catbox_9joc1v.png, 1MiB, 768x1152

Anonymous Wed 17 May 2023 11:56:44 No.7380311 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO 1683152862749420-890339753-sketc (...).png, 1MiB, 768x1152

Anonymous Wed 17 May 2023 11:59:42 No.7380312 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO 1684324769305740-2048903243-sket (...).png, 1MiB, 768x1152

Anonymous Wed 17 May 2023 12:04:02 No.7380315 Report

Quoted By:

this prototype guy loves his pink hair

Anonymous

View Same Google ImgOps iqdb SauceNAO 01108-2474414363.png, 2MiB, 1072x1608

Anonymous Wed 17 May 2023 12:16:02 No.7380329 Report

Quoted By:

futabie

Anonymous

View Same Google ImgOps iqdb SauceNAO 23669-2809700045-desert sorceres (...).png, 985KiB, 960x960

Anonymous Wed 17 May 2023 12:16:42 No.7380331 Report

Quoted By: >>7380336 >>7381089

>>7380126
>>7380127
>read this
>decide to try some old loras
>don't like them
>start rebaking
There's just no escape. Maybe I should just set up a bunch of wildcards and let the AI decide what I prompt instead

Anonymous

Anonymous Wed 17 May 2023 12:24:47 No.7380336 Report

Quoted By:

>>7380331
>sets up wildcards
>doesn't like the results every time
it's so owari

Anonymous

View Same Google ImgOps iqdb SauceNAO tmpt9he1ixc.png, 2MiB, 1024x1536

Anonymous Wed 17 May 2023 12:28:47 No.7380342 Report

Quoted By: >>7380345

Been messing around with SD, using image info and tips from these threads managed to learn some things.

Anonymous

Anonymous Wed 17 May 2023 12:29:27 No.7380345 Report

Quoted By: >>7380347

>>7380342
A blast from the past!

Anonymous

Anonymous Wed 17 May 2023 12:31:17 No.7380347 Report

Quoted By: >>7380349

>>7380345
Haha you recognize where I got the inspiration from?

Anonymous

Anonymous Wed 17 May 2023 12:32:46 No.7380349 Report

Quoted By: >>7380357

>>7380347
I mean more the skin and clothes texture, but what were you thinking about?

Anonymous

Anonymous Wed 17 May 2023 12:35:28 No.7380356 Report

Quoted By: >>7380634

>>7379300
>Auto tagger
new to this, what does auto tagger mean

Anonymous

Anonymous Wed 17 May 2023 12:35:46 No.7380357 Report

Quoted By: >>7380360

>>7380349
I'm new to this shit so I wouldn't know about that, I was referring to this: >>7373899

Anonymous

Anonymous Wed 17 May 2023 12:37:23 No.7380360 Report

Quoted By:

>change sampler
>was good in tests
>now i feel like everything's broken
owari da...
>>7380357
Ah. Welcome to the party, pal.

Anonymous

Anonymous Wed 17 May 2023 13:29:08 No.7380418 Report

Quoted By: >>7380429

>>7380242
>>7380231
actually can you even get a non-random outcome if you're using cos restarts?

Anonymous

Anonymous Wed 17 May 2023 13:32:29 No.7380422 Report

Quoted By: >>7380497

>>7380289
>link
damn didnt expect that to be the source
>img2img
did you consider using one of the controlnet models instead? there's so many by now one's oughta be better than img2img at this

Anonymous

View Same Google ImgOps iqdb SauceNAO 00007-305231912.png, 775KiB, 1024x1024

Anonymous Wed 17 May 2023 13:34:47 No.7380428 Report

Quoted By:

Anonymous

Anonymous Wed 17 May 2023 13:35:07 No.7380429 Report

Quoted By: >>7380644

>>7380418
iirc training has been deterministic due to some update a month ot 2 ago, as long as you don't have shuffle on? but that is easy to test just bake sonething twice

Anonymous

View Same Google ImgOps iqdb SauceNAO 1.png, 834KiB, 636x891

Anonymous Wed 17 May 2023 13:39:01 No.7380436 Report

Quoted By:

giratina ass hoe
sasuga testing grid

Anonymous

Anonymous Wed 17 May 2023 13:44:18 No.7380439 Report

Quoted By:

>>7379170
Welp, tested cu116-118 and turns out it was just a skill issue. Forgot I switched to trying out some retarded learning rates around the time I updated. Got spooked since this shit happened with embeds back in the day.

Anonymous

View Same Google ImgOps iqdb SauceNAO marisa kimono.png, 453KiB, 576x704

Anonymous Wed 17 May 2023 13:47:21 No.7380441 Report

Quoted By: >>7380444 >>7380463

Anonymous

View Same Google ImgOps iqdb SauceNAO marisa kimono.png, 1MiB, 1152x1408

Anonymous Wed 17 May 2023 13:50:39 No.7380444 Report

Quoted By: >>7380451 >>7380454

>>7380441
oh rip that was the wrong version.

Anonymous

Anonymous Wed 17 May 2023 13:53:33 No.7380451 Report

Quoted By:

>>7380444
it's nice

Anonymous

Anonymous Wed 17 May 2023 13:53:42 No.7380452 Report

Quoted By: >>7380456 >>7380457 >>7380459

>>7379837
No luck with this, can't figure out the right block in the Automatik1111 to put this in. Neither does modifying the actual webui-user.bat.

Anonymous

Anonymous Wed 17 May 2023 13:54:14 No.7380454 Report

Quoted By: >>7380463

>>7380444
catbox plx? cute style

Anonymous

Anonymous Wed 17 May 2023 13:55:52 No.7380456 Report

Quoted By:

>>7380452
Just do webui.bat (args here) when you go to launch it, no need to modify anything

Anonymous

View Same Google ImgOps iqdb SauceNAO file.png, 10KiB, 390x255

Anonymous Wed 17 May 2023 13:55:59 No.7380457 Report

Quoted By: >>7380465

>>7380452
Put it in the place where I have --xformers

Anonymous

Anonymous Wed 17 May 2023 13:58:01 No.7380459 Report

Quoted By: >>7380465

>>7380452
Use a non-shit vae instead of wasting your VRAM.

Anonymous

Anonymous Wed 17 May 2023 14:01:08 No.7380463 Report

Quoted By: >>7380466

>>7380441
>there's a 10 minute timeout to delete posts
man what is even the fucking point
>>7380454
not sure how useful it'll be for you since I'm testing a model mix I spun off last night on top of testing a lora revision while also, ALSO having a crazy base lora set on but here:
https://files.catbox.moe/wpqh6u.png
Model is a 7:3 based65:counterfeitv2.5 mix that's then mixed 8:2 with breakdomain m2000

Anonymous

View Same Google ImgOps iqdb SauceNAO paperPQKsmXZBq.png, 10KiB, 383x233

Anonymous Wed 17 May 2023 14:02:18 No.7380465 Report

Quoted By: >>7380467 >>7380468

>>7380457
Thanks, I had already done that and went through the entire restart machine gauntlet but no luck.

>>7380459
I'm using orangemix.vae.pt. I swear I did not have this kind of trouble until I decided to pull earlier today

Anonymous

Anonymous Wed 17 May 2023 14:02:26 No.7380466 Report

Quoted By:

>>7380463
Thanks
>schizo neg
based

Anonymous

Anonymous Wed 17 May 2023 14:03:24 No.7380467 Report

Quoted By: >>7380478 >>7380494

>>7380465
>I'm using orangemix.vae.pt
That's the shit one.

Anonymous

Anonymous Wed 17 May 2023 14:03:35 No.7380468 Report

Quoted By: >>7380478 >>7380494

>>7380465
orangemix vae is literally (in the actual literal meaning of the word) just the novelai VAE that has been renamed
Try using something else like https://huggingface.co/stabilityai/sd-vae-ft-mse-original/blob/main/vae-ft-mse-840000-ema-pruned.safetensors

Anonymous

View Same Google ImgOps iqdb SauceNAO tmpcardopt5.png, 687KiB, 1024x512

Anonymous Wed 17 May 2023 14:11:24 No.7380475 Report

Quoted By: >>7380484

More with the same model, lmk what you guys think or any suggestions (haven't looked too much into extensions yet and I'm using the NAI VAE, not sure if that's the way to go or not)

Anonymous

Anonymous Wed 17 May 2023 14:14:16 No.7380478 Report

Quoted By: >>7380484

>>7380467
>>7380468
Figures my stupid ass is also using some checkpoint that's not good either?

"AbyssOrangeMix2_nsfw_pruned_fp16_with_VAE.safetensors"

Anonymous

View Same Google ImgOps iqdb SauceNAO 01104-3731804198.png, 1MiB, 1048x1568

Anonymous Wed 17 May 2023 14:17:07 No.7380479 Report

Quoted By:

i have a slight problem

Anonymous

Anonymous Wed 17 May 2023 14:20:37 No.7380484 Report

Quoted By: >>7380494 >>7380519

>>7380475
Grab the one linked in the post above yours and do a side by side XYZ plot comparison in a with the novelai VAE and you can see the differences for yourself
Biggest downside of the novelai one is it sometimes randomly runs into memory issues for absolutely no reason

>>7380478
AOM is fine

Anonymous

Anonymous Wed 17 May 2023 14:23:25 No.7380489 Report

Quoted By: >>7380501

>>7379460
Is she pogging

Anonymous

Anonymous Wed 17 May 2023 14:27:14 No.7380494 Report

Quoted By:

>>7380467
>>7380468
Thanks. The other better vae seems to have stopped the random NaN spazzing.

>>7380484
>AOM is fine
Something went right for once, thanks

Anonymous

View Same Google ImgOps iqdb SauceNAO 00152-3934886101.png, 2MiB, 1448x1448

Anonymous Wed 17 May 2023 14:30:34 No.7380497 Report

Quoted By: >>7380539

>>7380422
i like that i can use any image as a source, i see it as txt2img with extra flavor i guess. the only downside is that it's more difficult to reproduce. feels bad when anons ask me for catbox and i can't really deliver. i should probably run the originals at low denoise and keep them so that i know where it comes from

>controlnet
i use it sometimes but i'm a vramlet, i can't use more than 2 models at once, on top of that it's slow. that's why i didn't feel the urge to update plus someone mentioned the newer version using more vram, idk how true that is but that's another reason i haven't tried updating. new models do look very interesting though, i liked seeing what anons did with them here. and dunno if it's going to be better than img2img for what i'm doing since img2img is much faster for me and even a small amount of tags is enough for it to result in something interesting. it can also create some hard to prompt poses/views more easily but that's still gacha

https://files.catbox.moe/w93k63.png
tried to get more meat with the new loras and messed around with an older prompt, it's pretty fleshy. i feel like i'm probably too used to the noise offset version since it makes red colors look intense. i'm gonna try this with other samplers later

Anonymous

Anonymous Wed 17 May 2023 14:33:02 No.7380501 Report

Quoted By:

>>7380489
human cock pogu

Anonymous

Anonymous Wed 17 May 2023 14:49:41 No.7380511 Report

Quoted By: >>7380513 >>7380514 >>7380516 >>7380523

>Download 371 images of a specific character
>After pruning pictures I find lacking in quality I'm left with 90 images

Every single time. Am I too picky, or are artists just too deranged? Is there any benefit to having ~200 mid-high quality images opposed to ~50-90 high-quality images?

Anonymous

Anonymous Wed 17 May 2023 14:51:03 No.7380513 Report

Quoted By: >>7380529

>>7380511
you might not be picky enough honestly

Anonymous

Anonymous Wed 17 May 2023 14:51:20 No.7380514 Report

Quoted By: >>7380529

>>7380511
Nope, that's completely normal
Every time I've finished deleting duplicates and low quality images I'm always left with only a tiny bit of the images I started with

Anonymous

Anonymous Wed 17 May 2023 14:53:47 No.7380516 Report

Quoted By: >>7380529

>>7380511
90's a lot.
I often end up with less after looking at over 2k images.

Anonymous

View Same Google ImgOps iqdb SauceNAO comparison.png, 1MiB, 1352x1140

Anonymous Wed 17 May 2023 14:54:57 No.7380519 Report

Quoted By:

>>7380484
Seems to work alright, not getting that NaN shit anymore.

Colors are more vibrant in the new one, only thing I don't like is the eyes, they look more blurry or less colorful or something I can't exactly put my finger on it.

Anonymous

Anonymous Wed 17 May 2023 14:55:36 No.7380520 Report

Quoted By:

Any tips on training style from hentai OVAs? I ran a small test and it didn't turn out good, should I nuke backgrounds and just leave single character?

Anonymous

Anonymous Wed 17 May 2023 14:58:52 No.7380522 Report

Quoted By: >>7380525

So does Sd-additional networks support lycoris/locon Loras out of the box now?

Anonymous

Anonymous Wed 17 May 2023 14:58:59 No.7380523 Report

Quoted By: >>7380529 >>7380542

>>7380511
>Is there any benefit to having ~200 mid-high quality images opposed to ~50-90 high-quality images?
yes

Anonymous

Anonymous Wed 17 May 2023 15:00:26 No.7380525 Report

Quoted By: >>7380527

>>7380522
it has for months

Anonymous

Anonymous Wed 17 May 2023 15:00:57 No.7380527 Report

Quoted By:

>>7380525
Nice

Anonymous

Anonymous Wed 17 May 2023 15:02:59 No.7380529 Report

Quoted By: >>7380531 >>7380542 >>7380551

>>7380513
>>7380514
>>7380516
Good, I was worried I am being too autistic. Might update the dataset after baking later. There are probably some suggestions as to which pictures to remove, even though that will probably be heavily opiniated

>>7380523
What are those? I'm guessing more flexibility?

Anonymous

Anonymous Wed 17 May 2023 15:04:25 No.7380531 Report

Quoted By: >>7380542

>>7380529
>What are those? I'm guessing more flexibility?
NTA, but I've generally tried to get even the shitter pics if my datasets are too small, because not having variety can be pretty catastrophic sometimes.

Anonymous

Anonymous Wed 17 May 2023 15:05:05 No.7380532 Report

Quoted By: >>7380535 >>7380536 >>7380650

>making character lora
>training with too few steps makes lora have great difficulty generating what the character should be consistently
>more steps improves accuracy but fries very quickly, can't find a sweet spot
>use the fried lora and fix up outputs to generate a ton of images to pad out the dataset
i feel dumb for not realizing i could just do this sooner

Anonymous

Anonymous Wed 17 May 2023 15:08:42 No.7380535 Report

Quoted By: >>7380538

>>7380532
How many images do you have? What's your LR?

Anonymous

Anonymous Wed 17 May 2023 15:09:04 No.7380536 Report

Quoted By: >>7380540

>>7380532
i'm nervous about data augmentation, it means you're baking in ai shittiness

Anonymous

Anonymous Wed 17 May 2023 15:09:54 No.7380538 Report

Quoted By: >>7380549 >>7380550 >>7380650

>>7380535
55 images right now, lr is i think 5e-4

Anonymous

Anonymous Wed 17 May 2023 15:10:05 No.7380539 Report

Quoted By:

>>7380497
>None
such a cool model actually. Seems the new one just doesnt draw the meatscape as well no matter what, hm

Anonymous

Anonymous Wed 17 May 2023 15:10:07 No.7380540 Report

Quoted By: >>7380541

>>7380536
Tried it once and it was gigashit.
Flipping, curiously, didn't do much.

Anonymous

Anonymous Wed 17 May 2023 15:10:54 No.7380541 Report

Quoted By:

>>7380540
i did some tests and flip helped a tiny bit so i just leave it on now

Anonymous

Anonymous Wed 17 May 2023 15:11:25 No.7380542 Report

Quoted By: >>7380543

>>7380531
>>7380529
>>7380523
The answer is "maybe". I think quality is more important, tiny datasets can work as long as the images are good. if you end up with 20 thats more than enough

Anonymous

Anonymous Wed 17 May 2023 15:12:57 No.7380543 Report

Quoted By: >>7380544

>>7380542
i'd agree but i still get pretty serious overfitting when i have less than 50 images total

Anonymous

Anonymous Wed 17 May 2023 15:13:37 No.7380544 Report

Quoted By: >>7380546

>>7380543
LR issue.

Anonymous

Anonymous Wed 17 May 2023 15:14:50 No.7380546 Report

Quoted By:

>>7380544
yeah i think that might be right all my testing on that front was old old old

Anonymous

Anonymous Wed 17 May 2023 15:15:40 No.7380549 Report

Quoted By: >>7380553

>>7380538
That seems pretty high, maybe try something lower like 1e-4 and see if that solves your problem

Anonymous

Anonymous Wed 17 May 2023 15:15:40 No.7380550 Report

Quoted By: >>7380553

>>7380538
5e-4 LR or unet LR? Remember if you set unet LR, LR itself is ignored

Anonymous

View Same Google ImgOps iqdb SauceNAO 00604-701738317.jpg, 392KiB, 1760x2080

Anonymous Wed 17 May 2023 15:16:30 No.7380551 Report

Quoted By: >>7380554

>>7380529
>What are those? I'm guessing more flexibility?
yeah basically. if you're using NAI it'll make it more compatible with other models and allow you to generate better looking outputs with them because a wider variety of worse looking artwork makes it less likely to learn style information.

Anonymous

Anonymous Wed 17 May 2023 15:16:52 No.7380553 Report

Quoted By: >>7380558

>>7380550
yeah it's unet lr

>>7380549
i still have 0 idea when to lower or raise LR desu but i'll try it out

Anonymous

Anonymous Wed 17 May 2023 15:17:08 No.7380554 Report

Quoted By:

>>7380551
>that image
jesus fuck

Anonymous

View Same Google ImgOps iqdb SauceNAO 00149-2762200492.png, 2MiB, 1264x1368

Anonymous Wed 17 May 2023 15:22:17 No.7380558 Report

Quoted By: >>7380560 >>7380562

>>7380553
In that case, I feel 5e-4 is a bit much, I use 2e-4 (or 0.0002125 in other cases) and it works fine at batch size 3, 5 and 7 for characters.

For conceps I use even lower values, but more repeats. (5e-5)

How much steps are you aiming with that LR?

Anonymous

Anonymous Wed 17 May 2023 15:24:18 No.7380560 Report

Quoted By: >>7380572 >>7380657 >>7380707 >>7380808

>>7380558
i'd never go above batch 2, it just degrades imo. i also train at really high dim so i use 5e-5 unet and 2e-5 text most of the time, and things are done at 1500 steps or less anyway, so i could probably get away with even lower lr and more steps

Anonymous

Anonymous Wed 17 May 2023 15:24:31 No.7380561 Report

Quoted By:

Alexi...

Anonymous

Anonymous Wed 17 May 2023 15:24:38 No.7380562 Report

Quoted By: >>7380572

>>7380558
usually between 1,100 and 1,900
i think maybe the reason i've not changed settings much is i got really lucky with one of my style trains

Anonymous

Anonymous Wed 17 May 2023 15:29:18 No.7380569 Report

Quoted By: >>7380593

>do a more extensive sampler gridtest
>the three top ones end up being the three that ranked top in previous test and that i used
huh
surprisingly comforting

Anonymous

View Same Google ImgOps iqdb SauceNAO 00159-772527211.png, 2MiB, 1264x1368

Anonymous Wed 17 May 2023 15:31:58 No.7380572 Report

Quoted By: >>7380581 >>7380586

>>7380560
I've tested BS 2 vs 3 and didn't notice much difference to be honest, and training at 1024x1024 (BS3) already takes too much. But the rest of your settings sound fine. I use 128 dim/alpha, and 5e-5 text lr. Has worked well for me so far. You're not the original anon right?

Oh and when I trained at 512x512, I used BS 12, and yeah that degrades vs BS 7 for example. But values lower than 5/4 seems fine.

>>7380562
Sounds fine, and probably high unet LR are pretty good for styles, but it can fry characters and concepts easily. (I use lower unet LR for concepts since I want to learn just the concept without changing the style).
What's your batch size?

Anonymous

Anonymous Wed 17 May 2023 15:35:43 No.7380581 Report

Quoted By: >>7380591

>>7380572
just 2
sometimes i test 3 or 4 on colab trains but i only have a 2060 so 2 is as high as i'll go local

Anonymous

Anonymous Wed 17 May 2023 15:39:24 No.7380586 Report

Quoted By: >>7380591

>>7380572
no you're takling with 3 or 4 people i think

Anonymous

View Same Google ImgOps iqdb SauceNAO 00174-370704044.png, 2MiB, 1264x1368

Anonymous Wed 17 May 2023 15:42:54 No.7380591 Report

Quoted By:

>>7380581
Good, then try with unet LR 2e-4, if you feel it doesn't learn everything you can try 2.5e-4 or 3e-4. If it still learns too much, try 1.5e-4 or 1e-4.

>>7380586
I hope I'm answering to the correct one this time IoI.

Anonymous

Anonymous Wed 17 May 2023 15:52:25 No.7380593 Report

Quoted By: >>7380595

>>7380569
which ones

Anonymous

View Same Google ImgOps iqdb SauceNAO 00000-120786-1girl, mature femal (...).png, 2MiB, 1280x1792

Anonymous Wed 17 May 2023 15:55:22 No.7380595 Report

Quoted By: >>7380603

>>7380593
Euler A, SDE Karras, DPM2 S a Karras.
Now my problem is that they're so close that I'd need to run like 30 seeds to establish a meaningful margin.

Anonymous

Anonymous Wed 17 May 2023 16:02:08 No.7380599 Report

Quoted By:

Sooo, AOM2_hard vs Based64, thoughts?

Anonymous

View Same Google ImgOps iqdb SauceNAO 5641546456.gif, 3MiB, 600x614

Anonymous Wed 17 May 2023 16:02:37 No.7380600 Report

Quoted By: >>7380608 >>7380625 >>7380633 >>7381079

Hemlo friends.
Someone knows how to do this?
I was thinking maybe using controlnet in each frame, but, I want to know if there's new AI that does this or what

Anonymous

View Same Google ImgOps iqdb SauceNAO 00196-179073907.png, 2MiB, 1264x1368

Anonymous Wed 17 May 2023 16:04:52 No.7380603 Report

Quoted By: >>7380611 >>7380710 >>7380722

>>7380595
Remember that DPM++ 2S a Karras and DPM++ SDE steps ar x2, so if you set 30 steps, in reality it is 60 steps.

You want to compare for example, 30 steps DPM++ 2S a Karras, DPM++ SDE vs 60 steps Euler a

Anonymous

Anonymous Wed 17 May 2023 16:05:41 No.7380608 Report

Quoted By:

>>7380600
https://rentry.org/AnimAnon

Anonymous

Anonymous Wed 17 May 2023 16:06:53 No.7380611 Report

Quoted By:

>>7380603
I wish there was normalization just for that.
But for now I'm testing the same amounts just to establish the samplers first.

Anonymous

Anonymous Wed 17 May 2023 16:19:02 No.7380625 Report

Quoted By:

>>7380600
demon slayer's art style never ceases to disappoint

Anonymous

Anonymous Wed 17 May 2023 16:25:30 No.7380633 Report

Quoted By:

>>7380600
The people doing these are training on spritesheets of animations and then compiling them back into an animated image, also why the res is usually so small

Anonymous

Anonymous Wed 17 May 2023 16:25:39 No.7380634 Report

Quoted By:

>>7380356

There's an extension where you upload a pic and it gives you what tags it thinks it should have, like reverse ingeneering an image.
You use it for mass tagging images of datasets to train lora.

Anonymous

Anonymous Wed 17 May 2023 16:27:09 No.7380637 Report

Quoted By:

>>7380123
Most of them

Anonymous

Anonymous Wed 17 May 2023 16:31:55 No.7380644 Report

Quoted By: >>7380645

>>7380429
update: no it's not deterministic with cos restarts embrace the gacha

Anonymous

Anonymous Wed 17 May 2023 16:33:37 No.7380645 Report

Quoted By: >>7380649

>>7380644
>all the people saying they don't change the seed for testing
>most of them using cosine with retards
it's over testbros

Anonymous

View Same Google ImgOps iqdb SauceNAO 00205-2003111286.png, 2MiB, 1024x1536

Anonymous Wed 17 May 2023 16:35:37 No.7380648 Report

Quoted By: >>7380722

>>7380206
The klf8 anime 2 vae

>>7380277
I'm not sure what that site is or does but eh I wasn't going to ever visit it anyway so go ahead and have fun. This is the only bj gen I made of her. I posted it a couple of threads back.

Anonymous

Anonymous Wed 17 May 2023 16:37:57 No.7380649 Report

Quoted By:

>>7380645
i don't think it really matters as long as you are sufficiently thorough in your testing, just need to add a second or third bake for testing

Anonymous

Anonymous Wed 17 May 2023 16:39:18 No.7380650 Report

Quoted By: >>7380651

>>7380532
>>7380538
retard. 55 images is more than enough, you're going to ruin your dataset with these AI gens.

Anonymous

Anonymous Wed 17 May 2023 16:39:48 No.7380651 Report

Quoted By:

>>7380650
good morning sir

Anonymous

Anonymous Wed 17 May 2023 16:43:10 No.7380657 Report

Quoted By: >>7380672

>>7380560
>5e5 to 2e5
weird values, isnt that text encode too high relatively to unet?
>never go above batch 2
why even stay at 2 and not go 1?

Anonymous

View Same Google ImgOps iqdb SauceNAO 35484746573465467465.jpg, 57KiB, 563x562

Anonymous Wed 17 May 2023 16:49:09 No.7380663 Report

Quoted By: >>7380665 >>7381076

man I just don't know how to use controlnet for inpainting those depth map hands, it just turns into a less shittier version of the regular inpaint blob hands

it probably does work as intended given how favorably people talk about it.. but I just have no fuckin clue settings wise my man

Anonymous

Anonymous Wed 17 May 2023 16:50:10 No.7380665 Report

Quoted By: >>7380666

>>7380663
catbox?

Anonymous

Anonymous Wed 17 May 2023 16:51:50 No.7380666 Report

Quoted By: >>7380668

>>7380665
it's a screencap my friend

Anonymous

Anonymous Wed 17 May 2023 16:52:58 No.7380668 Report

Quoted By: >>7380673

>>7380666
i don't believe you

Anonymous

Anonymous Wed 17 May 2023 16:58:34 No.7380672 Report

Quoted By: >>7380775

>>7380657
1 looks like shit, it doesn't learn concepts and overfits on things that appear in certain images. normally you do text = unet/2 and this is lower than that so not sure what you mean there

Anonymous

Anonymous Wed 17 May 2023 16:59:05 No.7380673 Report

Quoted By:

>>7380668
>good hands
clearly screencap

Anonymous

Anonymous Wed 17 May 2023 17:00:04 No.7380675 Report

Quoted By:

>yaoi lora and prompt outputting some gluck glucks peashooter 9000 extreme better than straight stuff

Anonymous

Anonymous Wed 17 May 2023 17:09:47 No.7380686 Report

Quoted By:

>>7380123
RTX 3060 Ti according to some article I read some weeks ago

Anonymous

View Same Google ImgOps iqdb SauceNAO 00065-2988791565.png, 1MiB, 1024x1536

Anonymous Wed 17 May 2023 17:11:57 No.7380688 Report

Quoted By: >>7380704

Holy shit this corrupted Saber looks scary

Anonymous

Anonymous Wed 17 May 2023 17:12:02 No.7380689 Report

Quoted By: >>7380706 >>7380709

I finally put together a mega
Enjoy my BS if you want.

https://mega.nz/folder/xDkWWRAZ#jrvHsIeim3Q1HyDUEHLjbw

Anonymous

Anonymous Wed 17 May 2023 17:23:28 No.7380704 Report

Quoted By:

>>7380688
this is just roboy-ky with tits

Anonymous

View Same Google ImgOps iqdb SauceNAO 00106-1116422342.png, 2MiB, 1024x1536

Anonymous Wed 17 May 2023 17:23:35 No.7380705 Report

Quoted By: >>7380722 >>7388582

Anonymous

Anonymous Wed 17 May 2023 17:23:46 No.7380706 Report

Quoted By: >>7380720

>>7380689
>Genshit
What's that trained on? The official character art images from twitter?

Anonymous

Anonymous Wed 17 May 2023 17:23:58 No.7380707 Report

Quoted By: >>7380712

>>7380560
what dim/alpha?

Anonymous

Anonymous Wed 17 May 2023 17:24:37 No.7380709 Report

Quoted By: >>7380720

>>7380689
>no preview images
>vague names
???

Anonymous

Anonymous Wed 17 May 2023 17:24:50 No.7380710 Report

Quoted By: >>7380903

>>7380603
Actually ended up back with Euler A by a big margin... I swear the gods are just messing with me. Shit's so all over the place.

Anonymous

Anonymous Wed 17 May 2023 17:25:57 No.7380712 Report

Quoted By: >>7380717 >>7380750

>>7380707
128/128dim locon (about equal to 256/256dim lora). i did notice i needed to increase lr for styles though, or maybe steps instead but i didnt try it

Anonymous

Anonymous Wed 17 May 2023 17:26:06 No.7380713 Report

Quoted By:

Is there a Chimosaku lora? Couldn't find one yet.

Anonymous

Anonymous Wed 17 May 2023 17:27:48 No.7380717 Report

Quoted By: >>7380718 >>7380719 >>7380729

>>7380712
>locon
why?

Anonymous

Anonymous Wed 17 May 2023 17:28:41 No.7380718 Report

Quoted By:

>>7380717
locondeez nuts
teehee

Anonymous

Anonymous Wed 17 May 2023 17:28:51 No.7380719 Report

Quoted By:

>>7380717
they work better

Anonymous

View Same Google ImgOps iqdb SauceNAO chrome_d9ibU0pw5S.png, 592KiB, 724x844

Anonymous Wed 17 May 2023 17:28:57 No.7380720 Report

Quoted By: >>7380726 >>7380731 >>7380733 >>7380735

>>7380706
the wiki has a lot of official character art
>>7380709
yeah yeah i'll try and update it when i have more time, this is mainly catharsis for me against that one anon who says we never share.

Anonymous

Anonymous Wed 17 May 2023 17:29:45 No.7380722 Report

Quoted By: >>7380743 >>7380903

>>7380603
>>7380648
>>7380705
catbox?

Anonymous

Anonymous Wed 17 May 2023 17:30:59 No.7380726 Report

Quoted By: >>7380746

>>7380720
Yeah, these are what I meant. What settings do you train your style loras on? Your Genshin and Fire Force LoRA seem pretty decent

Anonymous

Anonymous Wed 17 May 2023 17:31:51 No.7380729 Report

Quoted By:

>>7380717
objectively better for styles

Anonymous

Anonymous Wed 17 May 2023 17:32:14 No.7380731 Report

Quoted By: >>7380746

>>7380720
you should crop out the the logo

Anonymous

View Same Google ImgOps iqdb SauceNAO 00554-1461367833-masterpiece, si (...).png, 514KiB, 656x960

Anonymous Wed 17 May 2023 17:32:27 No.7380732 Report

Quoted By:

>>7379824

Anonymous

Anonymous Wed 17 May 2023 17:32:44 No.7380733 Report

Quoted By: >>7380746

>>7380720
lay off the jenshin impact

Anonymous

Anonymous Wed 17 May 2023 17:32:59 No.7380735 Report

Quoted By: >>7380739 >>7380741

>>7380720
Having preview images on mega is a disaster waiting to happen.

Anonymous

Anonymous Wed 17 May 2023 17:34:07 No.7380739 Report

Quoted By:

>>7380735
He could just bake one into the lora, but I don't like them either.

Anonymous

Anonymous Wed 17 May 2023 17:34:39 No.7380741 Report

Quoted By:

>>7380735
plenty of people have been doing it for months successfully. Just don't upload cradlesnatch shit like that alexitard

Anonymous

View Same Google ImgOps iqdb SauceNAO catbox_52eiip.png, 2MiB, 1024x1536

Anonymous Wed 17 May 2023 17:35:33 No.7380743 Report

Quoted By: >>7381454 >>7388582

>>7380722
I use inpainting to reroll gacha on hands, get rid of the extra digits and get rid of other errors.
https://files.catbox.moe/8vo3ga.png

https://files.catbox.moe/t2xddu.png

Anonymous

Anonymous Wed 17 May 2023 17:38:42 No.7380746 Report

Quoted By: >>7380752 >>7380765

>>7380731
>>7380733
you two have points

>>7380726
768 batch 4 (dadaptation defaults) 128/64 on screenshots from ffmpeg is my baseline for styles, i think you can check metadata somehow too

Anonymous

Anonymous Wed 17 May 2023 17:40:47 No.7380750 Report

Quoted By: >>7380754

>>7380712
cp decomposition?

Anonymous

Anonymous Wed 17 May 2023 17:41:41 No.7380752 Report

Quoted By: >>7380763

>>7380746
I'll have to name them better later and make previews for stuff that isn't obvious ("Whatever could fairy_tail.safetensors be?" isnt a question that needs answering.)

Anonymous

Anonymous Wed 17 May 2023 17:42:37 No.7380754 Report

Quoted By:

>>7380750
wtf is that

Anonymous

Anonymous Wed 17 May 2023 17:47:48 No.7380761 Report

Quoted By: >>7380878

can't seem to find cure black lora, but pretty much every other cure is available

Anonymous

Anonymous Wed 17 May 2023 17:48:18 No.7380763 Report

Quoted By: >>7380771

>>7380752
like what is "slime" for example

Anonymous

Anonymous Wed 17 May 2023 17:50:14 No.7380765 Report

Quoted By: >>7380771

>>7380746
What ffmpeg settings do you use? Do you prune bad screenshots/duplicates?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00006-511087163-1girl, blonde ha (...).png, 2MiB, 1280x1792

Anonymous Wed 17 May 2023 17:52:26 No.7380768 Report

Quoted By:

this is number two prostitute in kazakhstan
she's my sister :)

Anonymous

Anonymous Wed 17 May 2023 17:53:59 No.7380771 Report

Quoted By: >>7380773 >>7380784

>>7380763
"tensei shitara slime datta ken" is cumbersome
>>7380765
i've been experimenting with it so it varies but usually either keyframes or 4 screenshots/second (too many)

#4 screenshots every 1 seconds
ffmpeg -i "F:\Downloads\Kami\05.mkv" -vf fps=4/1 F:\Snips\Kami\1_2\$%05d.png

#keyframe only
ffmpeg -skip_frame nokey -i "$InputVideo" -vsync vfr -frame_pts true F:\Snips\Yugioh\$%05d.png

Anonymous

Anonymous Wed 17 May 2023 17:55:49 No.7380773 Report

Quoted By: >>7380777 >>7380787

>>7380771
>retarded
>seasonal haremshit watcher
many such cases

Anonymous

Anonymous Wed 17 May 2023 17:57:33 No.7380775 Report

Quoted By: >>7380851

>>7380672
should I run 512 batch 2 if I cant run batch 2 at 768 then?

Anonymous

Anonymous Wed 17 May 2023 17:57:44 No.7380777 Report

Quoted By: >>7380781 >>7380791

>>7380773
this but with anyone who watches anime

Anonymous

Anonymous Wed 17 May 2023 17:59:41 No.7380781 Report

Quoted By: >>7380785

>>7380777
you're so cool, want a cookie?

Anonymous

Anonymous Wed 17 May 2023 18:00:59 No.7380784 Report

Quoted By: >>7380795

>>7380771
So do you just throw everything into the dataset or do you look for good screenshots?

Anonymous

Anonymous Wed 17 May 2023 18:01:37 No.7380785 Report

Quoted By:

>>7380781
nah
i don't like them very much
got anything sour?

Anonymous

Anonymous Wed 17 May 2023 18:02:50 No.7380787 Report

Quoted By: >>7380792

>>7380773
“But I don’t want to be among retards," Alice remarked.
"Oh, you can’t help that," said the Cat: "we’re all retarded here. I’m retarded. You’re retarded."
"How do you know I’m retarded?" said Alice.
"You must be," said the Cat, "or you wouldn’t have come here.”

Anonymous

Anonymous Wed 17 May 2023 18:04:35 No.7380790 Report

Quoted By:

So anybody got those animation loras they were talking about?

Anonymous

Anonymous Wed 17 May 2023 18:04:40 No.7380791 Report

Quoted By: >>7380797

>>7380777
This but your momma's dried up vagina

Anonymous

Anonymous Wed 17 May 2023 18:05:25 No.7380792 Report

Quoted By:

>>7380787
>4chan is for retards
hang yourself retarded zoomer, its your degenerate kind propagating that shit and degrading this site

Anonymous

Anonymous Wed 17 May 2023 18:06:43 No.7380795 Report

Quoted By: >>7380799

>>7380784
it depends on the anime. for fairy tail I did but for others less so. keyframes works sometimes but not always and pruning 30k images is nightmarishly boring. I'll turn up the screenshot rate for episodes that have a lot of fanservice though, and maybe supplement with som stitched screenshots from gelbooru ir a guy on reddit called "LeonKevlar" who's a goldmine.

Anonymous

Anonymous Wed 17 May 2023 18:06:57 No.7380797 Report

Quoted By: >>7380801

>>7380791
it's only when she's around you
you have that effect on women, you know

Anonymous

Anonymous Wed 17 May 2023 18:08:34 No.7380799 Report

Quoted By: >>7380812

>>7380795
I just use hashing algos and tags to filter out images

Anonymous

Anonymous Wed 17 May 2023 18:09:09 No.7380800 Report

Quoted By: >>7380855 >>7380869

Should I watch megumins spinoff show?

Anonymous

Anonymous Wed 17 May 2023 18:09:12 No.7380801 Report

Quoted By: >>7380803

>>7380797
Of course you would know how your mom's vagina works, why wouldn't you

Anonymous

Anonymous Wed 17 May 2023 18:10:24 No.7380803 Report

Quoted By: >>7380820 >>7380844

>>7380801
your dad told me
the reason he never came back from the store is because he couldn't handle it around you anymore

Anonymous

Anonymous Wed 17 May 2023 18:11:31 No.7380808 Report

Quoted By: >>7380851

>>7380560
what loras/locons have u trained? got a link?

Anonymous

Anonymous Wed 17 May 2023 18:12:26 No.7380812 Report

Quoted By: >>7380859

>>7380799
do you have a recommended program for it? i figure its not a big deal to have repeats in a style.

something that would autogen pinups would be great too.

Anonymous

Anonymous Wed 17 May 2023 18:14:28 No.7380820 Report

Quoted By: >>7380823

>>7380803
Yeah dad discuss with you about your mom's vagina after his orgy with you and her the other day.

That post isn't the own you think it is lad....

Anonymous

View Same Google ImgOps iqdb SauceNAO 00047-3727021461-1girl, orange h (...).png, 2MiB, 1280x1792

Anonymous Wed 17 May 2023 18:14:33 No.7380821 Report

Quoted By:

notin better on a hot summer day
than some stroke inducing hay sex

Anonymous

Anonymous Wed 17 May 2023 18:15:44 No.7380823 Report

Quoted By: >>7380829

>>7380820
sad how that orgy took more time than he ever spent with you

Anonymous

View Same Google ImgOps iqdb SauceNAO 00129-1669590940.png, 2MiB, 1536x1024

Anonymous Wed 17 May 2023 18:16:02 No.7380824 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO 1684084846915217.jpg, 292KiB, 1447x1437

Anonymous Wed 17 May 2023 18:17:35 No.7380829 Report

Quoted By: >>7380834

>>7380823
And how would that be a problem, mom fucker

Anonymous

Anonymous Wed 17 May 2023 18:18:25 No.7380832 Report

Quoted By: >>7380839

>it was another retard larping as a trainer without a single lora to his name
man this thread's gone to shit

Anonymous

Anonymous Wed 17 May 2023 18:18:58 No.7380834 Report

Quoted By: >>7380837

>>7380829
>frog
damn i was having fun
never mind :(

Anonymous

Anonymous Wed 17 May 2023 18:20:30 No.7380837 Report

Quoted By:

>>7380834
>:(
Yeah it's around your bedtime, Johnny

Anonymous

Anonymous Wed 17 May 2023 18:21:11 No.7380839 Report

Quoted By: >>7380843

>>7380832
Who?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00488-3130020727-detailed pupils.png, 2MiB, 1024x1536

Anonymous Wed 17 May 2023 18:22:20 No.7380841 Report

Quoted By:

Anonymous

Anonymous Wed 17 May 2023 18:22:52 No.7380843 Report

Quoted By: >>7380850

>>7380839
read the thread to find ouit

Anonymous

Anonymous Wed 17 May 2023 18:24:00 No.7380844 Report

Quoted By: >>7380846

>>7380803
bro just admitted his mom got screwed by anons dad

Anonymous

Anonymous Wed 17 May 2023 18:25:19 No.7380846 Report

Quoted By: >>7380849

>>7380844
she had much fun
but it's soured now that she learnt that anon is a frogposter

Anonymous

View Same Google ImgOps iqdb SauceNAO 00001-497949186-1girl, orange ha (...).png, 2MiB, 1280x1792

Anonymous Wed 17 May 2023 18:27:19 No.7380848 Report

Quoted By: >>7380884

NEVA GIVE AP

Anonymous

View Same Google ImgOps iqdb SauceNAO 1684067444429319.png, 372KiB, 954x768

Anonymous Wed 17 May 2023 18:27:42 No.7380849 Report

Quoted By:

>>7380846
You're gonna be late for school tomorrow, boy. Actually this is a red board so you don't get to be here.

Anonymous

Anonymous Wed 17 May 2023 18:28:26 No.7380850 Report

Quoted By:

>>7380843
No.

Anonymous

Anonymous Wed 17 May 2023 18:28:28 No.7380851 Report

Quoted By: >>7380998

>>7380775
turn on gradient checkpointing homie
https://huggingface.co/docs/transformers/perf_train_gpu_one

>>7380808
https://mega.nz/folder/WjwmjSpA#_Sr_Gpl7WBXwCeSvV0eEtw

Anonymous

Anonymous Wed 17 May 2023 18:31:39 No.7380855 Report

Quoted By: >>7380861 >>7380863 >>7380868 >>7380886 >>7381309

>>7380800
>no aqua
what's the point

Anonymous

Anonymous Wed 17 May 2023 18:32:21 No.7380859 Report

Quoted By: >>7380862

>>7380812
>do you have a recommended program for it?
I'm doing this all in python , people call me dumb for doing all this by hand but I like the control.

Anonymous

Anonymous Wed 17 May 2023 18:32:52 No.7380861 Report

Quoted By: >>7380886

>>7380855
Almost cunny

Anonymous

Anonymous Wed 17 May 2023 18:33:58 No.7380862 Report

Quoted By: >>7380963

>>7380859
care to share scripts?

Anonymous

Anonymous Wed 17 May 2023 18:34:20 No.7380863 Report

Quoted By:

>>7380855
Virgin cunnygoober v. chad goddessman

Anonymous

Anonymous Wed 17 May 2023 18:36:05 No.7380868 Report

Quoted By: >>7380886

>>7380855
yun yun

Anonymous

Anonymous Wed 17 May 2023 18:36:10 No.7380869 Report

Quoted By:

>>7380800
the novel was good

Anonymous

Anonymous Wed 17 May 2023 18:41:18 No.7380878 Report

Quoted By:

>>7380761
anyone knows?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00016-3985764031-1girl, orange h (...).png, 2MiB, 1288x1800

Anonymous Wed 17 May 2023 18:46:07 No.7380884 Report

Quoted By: >>7380888

>>7380848
>the john deere rep came by after they learned she repaired the family tractor herself

Anonymous

Anonymous Wed 17 May 2023 18:46:18 No.7380886 Report

Quoted By: >>7380888 >>7380892 >>7380949

>>7380855
>>7380861
>>7380868
Oh hell I was posting in the wrong thread why you guys continue to homour this?

Anonymous

View Same Google ImgOps iqdb SauceNAO download.jpg, 263KiB, 1024x1024

Anonymous Wed 17 May 2023 18:48:14 No.7380888 Report

Quoted By:

>>7380884
love that artist
>>7380886
>wrong thread
baka anon

Anonymous

Anonymous Wed 17 May 2023 18:52:38 No.7380892 Report

Quoted By:

>>7380886
We're starved for contact.

Anonymous

Anonymous Wed 17 May 2023 18:55:59 No.7380900 Report

Quoted By: >>7380902

Google colab refugee here, doing a fresh install on my computer. Currently I have python 3.11 installed. Do I have to downgrade to python 3.10.6? I'm trying to install the easy lora training and it won't let me unless I have that version

> if version.major != 3 or version.minor != 10 or version.micro < 6:
> print("ERROR: You don't have python 3.10 installed, please install python 3.10, preferably 3.10.6, and add it to path")
> quit()

Anonymous

Anonymous Wed 17 May 2023 18:58:16 No.7380902 Report

Quoted By: >>7380909

>>7380900
You answered it yourself.
But you can just use conda if you want.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00256-3065510827.png, 2MiB, 1264x1368

Anonymous Wed 17 May 2023 18:58:47 No.7380903 Report

Quoted By: >>7380906

>>7380710
Interesting decision to go for Euler a, but if it's work for you then it's fine. Also, the top 3 you used do not converge, so they're pretty different each time.

My top 3 is possibly DPM++ SDE, DPM++ 2M SDE (installed using https://gist.github.com/crowsonkb/3ed16fba35c73ece7cf4b9a2095f2b78) and Euler a

>>7380722
https://files.catbox.moe/7hexiu.png

Anonymous

Anonymous Wed 17 May 2023 19:00:43 No.7380906 Report

Quoted By:

>>7380903
>Interesting decision to go for Euler a, but if it's work for you then it's fine
It's just weird to me that one time I'll masstest samplers and then come to a different conclusion altogether, but it is how it is with ML.

Anonymous

Anonymous Wed 17 May 2023 19:02:37 No.7380909 Report

Quoted By: >>7380919

>>7380902
So everyone just had an outdated version of python installed?

Anonymous

Anonymous Wed 17 May 2023 19:10:12 No.7380919 Report

Quoted By: >>7380925

>>7380909
Sure? It says to install it and most people didn't need it before.
My updated Linux is on 3.10 too.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00436-1175288016-nsfw, scan, (tr (...).png, 2MiB, 1024x1536

Anonymous Wed 17 May 2023 19:11:54 No.7380923 Report

Quoted By:

Anonymous

Anonymous Wed 17 May 2023 19:12:32 No.7380925 Report

Quoted By: >>7380930

>>7380919
Alright I guess I'll downgrade it then

Anonymous

Anonymous Wed 17 May 2023 19:14:57 No.7380930 Report

Quoted By: >>7380981

>>7380925
Or, like I said, just use conda.
You shouldn't really be installing packages on system python anyway, but it seems like the installers do that?

Anonymous

Anonymous Wed 17 May 2023 19:30:24 No.7380949 Report

Quoted By:

>>7380886
based retard
please stay

Anonymous

View Same Google ImgOps iqdb SauceNAO 00273-1933563816.png, 2MiB, 1264x1368

Anonymous Wed 17 May 2023 19:39:18 No.7380959 Report

Quoted By: >>7380962 >>7380964 >>7380965 >>7380966

why is she mad

Anonymous

Anonymous Wed 17 May 2023 19:40:36 No.7380962 Report

Quoted By:

>>7380959
no room to work

Anonymous

Anonymous Wed 17 May 2023 19:40:57 No.7380963 Report

Quoted By:

>>7380862
I'm at work so I'll need to after I get home

Anonymous

Anonymous Wed 17 May 2023 19:41:05 No.7380964 Report

Quoted By:

>>7380959
its hot and the lora still aint coming out right

Anonymous

Anonymous Wed 17 May 2023 19:42:22 No.7380965 Report

Quoted By:

>>7380959
Lost pity on Ganyu and got Qiqi'd again.

Anonymous

Anonymous Wed 17 May 2023 19:43:33 No.7380966 Report

Quoted By:

>>7380959
she's browsing hdg

Anonymous

Anonymous Wed 17 May 2023 19:58:15 No.7380981 Report

Quoted By:

>>7380930
Apparently conda doesn't support 3.11 yet.

Anonymous

Anonymous Wed 17 May 2023 20:12:56 No.7380997 Report

Quoted By: >>7381006 >>7381158

lmao based66 is kinda garbage tbdesu

Anonymous

Anonymous Wed 17 May 2023 20:14:11 No.7380998 Report

Quoted By: >>7381001

>>7380851
Pretty mid settings.
https://files.catbox.moe/m4vuy4.png
https://files.catbox.moe/yv93v0.png
https://files.catbox.moe/wylya5.png
at least it sparked the idea for this unholy mix https://files.catbox.moe/a552jp.png
pretty sure LRs are too low to grasp character details right, as you mentioned in one of the posts - baking it longer will most likely just overfit it without improvement.

Anonymous

Anonymous Wed 17 May 2023 20:15:31 No.7381001 Report

Quoted By: >>7381011

>>7380998
yeah i don't suggest these for characters or styles didn't seem to work at all the couple times i tried

Anonymous

Anonymous Wed 17 May 2023 20:17:05 No.7381006 Report

Quoted By: >>7381008 >>7381266

>>7380997
65 was too, so I'm seeing a trend.
Too bad AOM did the same thing.

Anonymous

Anonymous Wed 17 May 2023 20:18:37 No.7381008 Report

Quoted By: >>7381013 >>7381014

>>7381006
>I'm seeing a trend.
What's the trend?

Anonymous

Anonymous Wed 17 May 2023 20:19:26 No.7381011 Report

Quoted By: >>7381012

>>7381001
so why would you use a locon if you're not training a style? isnt this the worst lora "type" to train concepts on?

Anonymous

Anonymous Wed 17 May 2023 20:19:53 No.7381012 Report

Quoted By:

>>7381011
no

Anonymous

Anonymous Wed 17 May 2023 20:20:24 No.7381013 Report

Quoted By:

>>7381008
people gilding the lily

Anonymous

Anonymous Wed 17 May 2023 20:21:01 No.7381014 Report

Quoted By:

>>7381008
Each version being more terrible than the last.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00322-4141419164.png, 2MiB, 1264x1368

Anonymous Wed 17 May 2023 20:22:13 No.7381016 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO 00331-131937030.png, 2MiB, 1264x1368

Anonymous Wed 17 May 2023 20:28:58 No.7381025 Report

Quoted By: >>7381033

Anonymous

Anonymous Wed 17 May 2023 20:32:18 No.7381029 Report

Quoted By:

>>7379521
These breasts are amazing.

Anonymous

Anonymous Wed 17 May 2023 20:34:40 No.7381033 Report

Quoted By: >>7381052

>>7381025
I hate the lycoris schizo, yurifags are mental

Anonymous

Anonymous Wed 17 May 2023 20:46:12 No.7381052 Report

Quoted By: >>7381058

>>7381033
who is he talking to

Anonymous

View Same Google ImgOps iqdb SauceNAO 1684324769306153-1011002401-nsfw (...).png, 949KiB, 768x1152

Anonymous Wed 17 May 2023 20:48:37 No.7381055 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO 1684324769306150-1011002401-nsfw (...).png, 1MiB, 768x1152

Anonymous Wed 17 May 2023 20:48:54 No.7381056 Report

Quoted By:

Anonymous

Anonymous Wed 17 May 2023 20:49:54 No.7381058 Report

Quoted By:

>>7381052
https://myanimelist.net/character/204620/Takina_Inoue its a lycoris lora disguised as a model

Anonymous

Anonymous Wed 17 May 2023 20:53:47 No.7381063 Report

Quoted By:

What if loras gained sentience?

Anonymous

Anonymous Wed 17 May 2023 20:57:33 No.7381067 Report

Quoted By:

don't care, still calling then locons

Anonymous

View Same Google ImgOps iqdb SauceNAO 00007-1953848051-1girl, green sk (...).png, 2MiB, 1280x1792

Anonymous Wed 17 May 2023 21:00:22 No.7381070 Report

Quoted By:

Anonymous

Anonymous Wed 17 May 2023 21:03:57 No.7381076 Report

Quoted By:

>>7380663
How do you use those depth hands anyways, all the yt vids are kinda shit and don't explain the updated controlnet/models

Anonymous

Anonymous Wed 17 May 2023 21:06:21 No.7381079 Report

Quoted By:

>>7380600
people get barely hands decent

Anonymous

View Same Google ImgOps iqdb SauceNAO 00012-2157715182-1girl, black ha (...).png, 2MiB, 1280x1792

Anonymous Wed 17 May 2023 21:08:49 No.7381086 Report

Quoted By: >>7381111

Anonymous

View Same Google ImgOps iqdb SauceNAO 00434-3334802266.png, 2MiB, 1024x1536

Anonymous Wed 17 May 2023 21:09:40 No.7381089 Report

Quoted By: >>7381313

>>7380331
Dark Souls II Desert Sorceress? Is there a LoRA?

Anonymous

Anonymous Wed 17 May 2023 21:20:03 No.7381100 Report

Quoted By: >>7381104

I've been out the loop since like october.
Is that shitty gook sameface I see in every other picture across entire intenet caused by model or loras?

Anonymous

Anonymous Wed 17 May 2023 21:20:41 No.7381104 Report

Quoted By: >>7381106

>>7381100
Models, mostly.
Japs injected them with so much pedoshit it all melted together.

Anonymous

Anonymous Wed 17 May 2023 21:21:52 No.7381106 Report

Quoted By: >>7381108 >>7381110

>>7381104
Oh dang it I hope I can find some good model then

Anonymous

Anonymous Wed 17 May 2023 21:22:47 No.7381108 Report

Quoted By:

>>7381106
Oh yeah, no problem. The good ones are separated well.

Anonymous

Anonymous Wed 17 May 2023 21:23:46 No.7381110 Report

Quoted By:

>>7381106
I've had a lot of success with AcidBurn and Based65 (I don't like Based66 as much). I'm sure other people have their own recommendations, but I think those are both good for general purpose anime.

Anonymous

Anonymous Wed 17 May 2023 21:25:15 No.7381111 Report

Quoted By: >>7381114 >>7381138 >>7381145

>>7381086
Catbox for this?
For some weird reason, I've never been able to get this pose. Dunno why, there are just some things that don't seem to prompt well for me.

Anonymous

Anonymous Wed 17 May 2023 21:26:30 No.7381114 Report

Quoted By: >>7381147

>>7381111
Unfortunately for you there is no pose data in this. I'll get it to you in a second tho.

Anonymous

Anonymous Wed 17 May 2023 21:28:37 No.7381116 Report

Quoted By:

I was wondering but can updating webui using git pull when being like few weeks behind fuck up the settings compared to fresh install or is there nothing to worry about?
I don't really get how git works

Anonymous

View Same Google ImgOps iqdb SauceNAO 00722-4056918342.png, 2MiB, 1024x1536

Anonymous Wed 17 May 2023 21:29:11 No.7381117 Report

Quoted By: >>7381119 >>7381783

Shorstack lora on civit. I can't take it bros.

gobbo girls forever.

Anonymous

Anonymous Wed 17 May 2023 21:29:59 No.7381119 Report

Quoted By:

>>7381117
nothing gobbo about this besides green

Anonymous

Anonymous Wed 17 May 2023 21:37:36 No.7381133 Report

Quoted By:

>>7379831
I used the headpat lora from civitai, a lot of inpaint, and photoshop,

the penises are photoshopped first, partially hand drawn on top and then inpainted

Anonymous

Anonymous Wed 17 May 2023 21:38:36 No.7381138 Report

Quoted By: >>7381147

>>7381111
https://files.catbox.moe/yd9zro.png

Anonymous

Anonymous Wed 17 May 2023 21:43:36 No.7381145 Report

Quoted By: >>7381155

>>7381111
"wariza, legs up, knees together, pussy, pussy peek, facing viewer, sitting" may help you

Anonymous

Anonymous Wed 17 May 2023 21:46:34 No.7381147 Report

Quoted By:

>>7381114
>>7381138
Ah, might just be luck of the draw. I appreciate it nonetheless. I'll play around with it and see if I can learn something.

Anonymous

Anonymous Wed 17 May 2023 21:49:53 No.7381155 Report

Quoted By:

>>7381145
I admit, I hadn't considered using wariza that way. Might be just what I need.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00011-2227279161-best quality, 1 (...).png, 3MiB, 1536x2304

Anonymous Wed 17 May 2023 21:51:08 No.7381158 Report

Quoted By: >>7381240 >>7381269

>>7380997
What's wrong with it?
(Forgive the sloppy gen, I usually just lurk but thought I should at least post a picture if I'm going to ask a question.)

Anonymous

View Same Google ImgOps iqdb SauceNAO 20230509195512-3358090102-master (...).png, 2MiB, 1024x1536

Anonymous Wed 17 May 2023 22:25:58 No.7381207 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO 1684324769306159-3690327404-nsfw (...).png, 937KiB, 768x1152

Anonymous Wed 17 May 2023 22:26:56 No.7381208 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO catbox_ei1wwu.png, 1MiB, 1024x1280

Anonymous Wed 17 May 2023 22:39:33 No.7381235 Report

Quoted By:

Anonymous

Anonymous Wed 17 May 2023 22:42:27 No.7381240 Report

Quoted By:

>>7381158
NTA, but it's just fried and not better than 64.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00093.jpg, 327KiB, 1408x2048

Anonymous Wed 17 May 2023 22:49:05 No.7381247 Report

Quoted By: >>7381249 >>7381254

Did i dun a gud?

Anonymous

Anonymous Wed 17 May 2023 22:50:03 No.7381249 Report

Quoted By: >>7381259

>>7381247
i mean
not really but it's a fun picture

Anonymous

View Same Google ImgOps iqdb SauceNAO 00098.png, 4MiB, 1264x2048

Anonymous Wed 17 May 2023 22:55:07 No.7381254 Report

Quoted By:

>>7381247
b-but you said you wouldn't cum in me Goku

Anonymous

Anonymous Wed 17 May 2023 22:56:12 No.7381259 Report

Quoted By: >>7381262

>>7381249
Wat do improve?

Anonymous

Anonymous Wed 17 May 2023 22:57:41 No.7381262 Report

Quoted By:

>>7381259
Hands are FUBAR, pupils nonexistent, and it's super blurry.
You'd have to give gatobox, but it's probably a style lora if you're using one.
Otherwise pretty good.

Anonymous

Anonymous Wed 17 May 2023 23:00:44 No.7381266 Report

Quoted By: >>7381269 >>7381273

>>7381006
I don't think it's fried since there's little artifacting at CFGs like 7-11, but style LoRA compatibility is abysmal for some reason.
I would likely have loved it otherwise, since it doesn't have much 3D influence.

Anonymous

Anonymous Wed 17 May 2023 23:01:51 No.7381269 Report

Quoted By:

>>7381266 is meant for >>7381158

Anonymous

View Same Google ImgOps iqdb SauceNAO 00337-432422760.png, 2MiB, 1264x1368

Anonymous Wed 17 May 2023 23:06:27 No.7381273 Report

Quoted By: >>7381323

>>7381266
NTA, but when you say it is not good with LoRAs, do you mean getting fried results? I remember for example that

B64v3 = good with LoRAs
B65= kinda bad with LoRAs (easily getting fried images)

AOM2 = good with LoRAs
AOM3 = same case as B65

I'm not sure how that happens, more models in the "mix" add more possible collisions between the tokens?

Anonymous

Anonymous Wed 17 May 2023 23:14:02 No.7381286 Report

Quoted By: >>7381290 >>7381294

Learning to train LoRAs right now, when using the kohya scripts, I've noticed that most of the posts on rentry prefix the settings with a $, from reading the actual README on the repo though, it seems I should be using a .json file with the desired settings I want, is this a case of the posts being outdated or are these two separate things entirely?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00353-181509779.png, 2MiB, 1280x1408

Anonymous Wed 17 May 2023 23:16:01 No.7381290 Report

Quoted By: >>7381295

>>7381286
kohya scripts that are written in .ps1 format, use $ to define a variable. json syntax should be another one I think (I use .ps1 scripts)

Anonymous

Anonymous Wed 17 May 2023 23:20:13 No.7381294 Report

Quoted By: >>7381299 >>7381305

>>7381286
Kohya's readme assumes you use the web GUI instead where you can save/import your settings as .json while I believe the rentry that gets passed around uses a different file to train without needing to open the GUI in your browser

Anonymous

Anonymous Wed 17 May 2023 23:20:42 No.7381295 Report

Quoted By: >>7381296

>>7381290
Alright, thanks for clearing that up, I don't see any powershell scripts in my install directory, did you create them yourself and adjust the variables before manually running the python script?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00369-2513026845.png, 2MiB, 1280x1408

Anonymous Wed 17 May 2023 23:23:36 No.7381296 Report

Quoted By: >>7381299

>>7381295
Someone when lora just started to exist, created a .ps1 script and then have been modying that to my needs. Also, I can automate "sorta easily" to queue lora training.

Here is an example of what I use

https://files.catbox.moe/f50q80.ps1

You can execute that in the kohya ss scripts folder, or add a cd on the start of the file to go to that directory.
"cd your/kohyass/directory"

Anonymous

Anonymous Wed 17 May 2023 23:26:11 No.7381299 Report

Quoted By:

>>7381296
Ok I get it now, thanks.
>>7381294
Are you talking about this: https://github.com/ddPn08/kohya-sd-scripts-web0yavui?

Anonymous

Anonymous Wed 17 May 2023 23:35:53 No.7381305 Report

Quoted By: >>7381308

>>7381294
Easyscripts also use jsons, but I think they're different than the GUI ones.

Anonymous

Anonymous Wed 17 May 2023 23:38:46 No.7381308 Report

Quoted By: >>7381310

>>7381305
the jsons are optional for ezscripts you can edit the training script kinda like the kohya one

Anonymous

Anonymous Wed 17 May 2023 23:39:48 No.7381309 Report

Quoted By:

>>7380855
god so based GOD SO BASED GGOOODDD SOOO FUCKING BASSSEEDDD

Anonymous

Anonymous Wed 17 May 2023 23:40:03 No.7381310 Report

Quoted By:

>>7381308
True, but the jsons are pretty convenient. I've had people complain they don't work on guis before though.

Anonymous

Anonymous Wed 17 May 2023 23:44:21 No.7381313 Report

Quoted By:

>>7381089
https://civitai.com/models/62188/desert-sorceress-dark-souls-2-outfit

Anonymous

Anonymous Wed 17 May 2023 23:50:37 No.7381322 Report

Quoted By:

>lora actually understands reptilian feet
neat

Anonymous

Anonymous Wed 17 May 2023 23:50:37 No.7381323 Report

Quoted By:

>>7381273
No, in this case I mean that it literally does not work well, i.e. the outputs are not enough in the artist's style unless the LoRA is way overfit, the weight is raised, or it's very similar to the model's default style anyway. LoCons work a bit better but it shouldn't be necessary to rebake LoRAs that work fine on other models. Currently at work but I'll post an XYZ later if I remember.

Anonymous

Anonymous Wed 17 May 2023 23:59:24 No.7381333 Report

Quoted By: >>7381349

https://arxiv.org/abs/2305.08891

Anonymous

View Same Google ImgOps iqdb SauceNAO 00075-3392967156.png, 1MiB, 1024x1280

Anonymous Thu 18 May 2023 00:07:24 No.7381344 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO 00409-3788218448.png, 2MiB, 1280x1408

Anonymous Thu 18 May 2023 00:08:14 No.7381345 Report

Quoted By: >>7381369 >>7381392 >>7381466

Damn I love buff waifus

Anonymous

Anonymous Thu 18 May 2023 00:10:24 No.7381349 Report

Quoted By:

>>7381333
oh hey more ways to deal with the noise issue cool

Anonymous

View Same Google ImgOps iqdb SauceNAO 00002-3752154593.png, 570KiB, 768x512

Anonymous Thu 18 May 2023 00:14:23 No.7381353 Report

Quoted By: >>7381766

Anonymous

View Same Google ImgOps iqdb SauceNAO 00022-500534983-1girl, purple ha (...).png, 2MiB, 1280x1792

Anonymous Thu 18 May 2023 00:22:32 No.7381359 Report

Quoted By: >>7381378

lol

Anonymous

View Same Google ImgOps iqdb SauceNAO 00146.png, 3MiB, 1344x2016

Anonymous Thu 18 May 2023 00:24:34 No.7381363 Report

Quoted By: >>7381487

Anonymous

Anonymous Thu 18 May 2023 00:34:40 No.7381369 Report

Quoted By:

>>7381345
this model sucks

Anonymous

View Same Google ImgOps iqdb SauceNAO 00453-1923602024.png, 2MiB, 1280x1536

Anonymous Thu 18 May 2023 00:40:49 No.7381376 Report

Quoted By: >>7381396 >>7381983

Does someone knows if there's a "negative" lora like easynegative, deepverybadneg, etc.

If not, I could try training one to see how it goes.

At least in the reverse (add detail lora https://civitai.com/models/58390) somehow it works really good

Anonymous

View Same Google ImgOps iqdb SauceNAO 00006-2237679403-1girl, purple h (...).png, 3MiB, 1280x1792

Anonymous Thu 18 May 2023 00:43:47 No.7381378 Report

Quoted By: >>7381674

>>7381359
Another try.

Anonymous

Anonymous Thu 18 May 2023 00:43:48 No.7381379 Report

Quoted By: >>7381398 >>7381400 >>7381403 >>7381408

so testing out a new workflow ,
anyone character loras that no ones done , or whos current loras are left something to be desired?

Anonymous

Anonymous Thu 18 May 2023 00:44:27 No.7381381 Report

Quoted By: >>7381382 >>7381383 >>7381385 >>7381402

gimme prompts to gen the most fucked up hands

Anonymous

Anonymous Thu 18 May 2023 00:45:22 No.7381382 Report

Quoted By:

>>7381381
(scary movie 2, caretaker, strong hand:1.5)

Anonymous

Anonymous Thu 18 May 2023 00:45:35 No.7381383 Report

Quoted By:

>>7381381
(many fingers:6)

Anonymous

Anonymous Thu 18 May 2023 00:46:07 No.7381385 Report

Quoted By:

>>7381381
hand

Anonymous

Anonymous Thu 18 May 2023 00:49:58 No.7381392 Report

Quoted By:

>>7381345
damn I hate lycoris recoil

Anonymous

View Same Google ImgOps iqdb SauceNAO 00145.png, 2MiB, 1680x1120

Anonymous Thu 18 May 2023 00:51:42 No.7381396 Report

Quoted By: >>7381410

>>7381376
like the style anon, have a catbox?

Anonymous

Anonymous Thu 18 May 2023 00:52:46 No.7381398 Report

Quoted By: >>7381422

>>7381379
You could try C.C.
There's LoRAs but none can do even one of her outfits accurately or consistently.

Anonymous

View Same Google ImgOps iqdb SauceNAO 1684324769306174-1610640430-full (...).png, 1MiB, 768x1152

Anonymous Thu 18 May 2023 00:53:24 No.7381399 Report

Quoted By:

Anonymous

Anonymous Thu 18 May 2023 00:53:56 No.7381400 Report

Quoted By:

>>7381379
there are like a zillion characters that have either shit or nothing

Anonymous

Anonymous Thu 18 May 2023 00:54:29 No.7381402 Report

Quoted By:

>>7381381
holding

Anonymous

Anonymous Thu 18 May 2023 00:55:51 No.7381403 Report

Quoted By:

>>7381379
basically no characters have good loras lmao

Anonymous

Anonymous Thu 18 May 2023 00:58:05 No.7381408 Report

Quoted By:

>>7381379
aqua
not because it's bad but because I want more aqua

Anonymous

View Same Google ImgOps iqdb SauceNAO 00471-4287706577.png, 2MiB, 1280x1536

Anonymous Thu 18 May 2023 00:58:19 No.7381410 Report

Quoted By: >>7381420

>>7381396
https://files.catbox.moe/x235s4.png

Anonymous

View Same Google ImgOps iqdb SauceNAO 00144.png, 2MiB, 1120x1680

Anonymous Thu 18 May 2023 01:03:44 No.7381420 Report

Quoted By: >>7381429

>>7381410
Thanks! What's the model you used? I can't find it on civitai with the hash

Anonymous

Anonymous Thu 18 May 2023 01:05:20 No.7381422 Report

Quoted By: >>7381437

>>7381398
>C.C.
from code geass?

Anonymous

Anonymous Thu 18 May 2023 01:07:07 No.7381427 Report

Quoted By:

can anyone share a catbox image sample for alexi's gpqbao lora? i want the image for a preview for my lora as well as a sample prompt i can copy+paste when genning

Anonymous

View Same Google ImgOps iqdb SauceNAO 00478-1340913397.png, 2MiB, 1280x1536

Anonymous Thu 18 May 2023 01:08:23 No.7381429 Report

Quoted By: >>7381430

>>7381420
I did it myself, basically various models mixed and then 50% of xxxmix.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00084.png, 3MiB, 1344x2016

Anonymous Thu 18 May 2023 01:10:20 No.7381430 Report

Quoted By: >>7381516

>>7381429
Ooh, do you think you could provide a download perhaps..?

Anonymous

Anonymous Thu 18 May 2023 01:15:46 No.7381434 Report

Quoted By: >>7381439

i've been messing around with ai casanova's suggested method of using dim from weights and it's actually working better than i thought it would

you can train a LORA on a broad dataset with shittier, innacurate images in it, reduce its dimensions, and then train it again on nothing but super high quality/very accurate images with more or less just the contents of the LORA changing, rather than the art style. I tried this by training a base lora on 2d fanart and a second pass on nothing but 3d images and the output style didn't change much, small details just became more accurate.

Anonymous

Anonymous Thu 18 May 2023 01:20:57 No.7381437 Report

Quoted By: >>7381440

>>7381422
Yes.

Anonymous

Anonymous Thu 18 May 2023 01:22:20 No.7381439 Report

Quoted By: >>7381442

>>7381434
>ai casanova's suggested method
?

Anonymous

Anonymous Thu 18 May 2023 01:22:53 No.7381440 Report

Quoted By: >>7381450

>>7381437
can you give me examples of it missing the mark in so far as itfits like you said?

Anonymous

Anonymous Thu 18 May 2023 01:24:09 No.7381442 Report

Quoted By:

>>7381439
https://github.com/kohya-ss/sd-scripts/pull/491

Anonymous

Anonymous Thu 18 May 2023 01:33:21 No.7381450 Report

Quoted By: >>7381461

>>7381440
well there's this one which is completely generic and can't do any outfits at all
https://civitai.com/models/28439/cc-code-geass-lelouch-of-the-rebellion-or-locon
and then there's this one that can sort of do her most basic outfit but as you can see from the examples, you have to raise the tokens' weight to 1.4 or 1.6 and it still doesn't get the outfit right.
https://civitai.com/models/6358/tsumasaky-cc-code-geass

Anonymous

Anonymous Thu 18 May 2023 01:36:44 No.7381454 Report

Quoted By: >>7381476

>>7380743
Love your work anon but yooo is the inCase lora you use the one off Civitai just renamed, or a different one entirely? Cus I tried running the BJ prompt from catbox using the current lora on Civ info and all I got back was a lot of futa lmao

Anonymous

View Same Google ImgOps iqdb SauceNAO tmpmf1d3yfg.png, 991KiB, 1024x1024

Anonymous Thu 18 May 2023 01:38:24 No.7381455 Report

Quoted By: >>7381538

Anonymous

Anonymous Thu 18 May 2023 01:48:47 No.7381461 Report

Quoted By: >>7381511

>>7381450
those looks fine to me

Anonymous

View Same Google ImgOps iqdb SauceNAO The girl on the left agrees.jpg, 1MiB, 2304x2816

Anonymous Thu 18 May 2023 01:53:35 No.7381466 Report

Quoted By:

>>7381345

That they are.

Anonymous

Anonymous Thu 18 May 2023 02:00:47 No.7381476 Report

Quoted By: >>7381494 >>7381732

>>7381454
>Love your work anon
ty ty
>yooo is the inCase lora you use the one off Civitai just renamed, or a different one entirely?
No I do not use the incase LoRA from civit.
I have 3 incase LoRAs. All of them are done by anon who uploaded 2 of them to gayshit. The first one was trained on 512 x 512 data, the second was trained on 768x 768 data. The later one wasn't to anon's (or my) tastes so anon is currently in the process of baking a 3rd but it's a WIP.
Here is a link to v1 of the incase LoRA trained on 512 x 512 data: https://files.catbox.moe/ehvb4v.safetensors
Here is the WIP for the latest version: https://files.catbox.moe/pr9gyg.safetensors
And the 2nd version is currently on gayshit. I don't use this one but if it's something you're interested in using I'm just letting you know.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00661-186929224-masterpiece, ((( (...).png, 725KiB, 768x1120

Anonymous Thu 18 May 2023 02:04:52 No.7381480 Report

Quoted By:

Anonymous

Anonymous Thu 18 May 2023 02:09:57 No.7381487 Report

Quoted By:

>>7381363
beautiful anon

Anonymous

View Same Google ImgOps iqdb SauceNAO 00053-3070823720.png, 907KiB, 768x1152

Anonymous Thu 18 May 2023 02:09:57 No.7381488 Report

Quoted By: >>7381504

An accidental gen when I forgot to add character LoRA

Anonymous

Anonymous Thu 18 May 2023 02:14:13 No.7381494 Report

Quoted By: >>7382058 >>7382082

>>7381476
man you're just a baller in every sense of the word. sharing prompts, sharing great art, sharing resources. TY bro, you're a real one

Anonymous

Anonymous Thu 18 May 2023 02:18:12 No.7381504 Report

Quoted By: >>7381520

>>7381488
holy fuck anon just give a mega or catbox of that model

Anonymous

Anonymous Thu 18 May 2023 02:21:58 No.7381511 Report

Quoted By: >>7381521

>>7381461
lol, if you don't wanna do it that's fine; it was just a suggestion since you (assuming you're that anon) asked and I can train it myself later.
That said, being able to gen only the face, which even the base model can do without any LoRAs is definitely not "fine".

Anonymous

View Same Google ImgOps iqdb SauceNAO 00515-2271575306.png, 2MiB, 1280x1536

Anonymous Thu 18 May 2023 02:23:22 No.7381516 Report

Quoted By: >>7381526

>>7381430
Maybe I will upload to HF (anon mixer here), not sure if they will let me upload things after the AOM2 mixes IoI

Anonymous

View Same Google ImgOps iqdb SauceNAO 00001-2887475754.png, 914KiB, 768x1152

Anonymous Thu 18 May 2023 02:26:56 No.7381520 Report

Quoted By:

>>7381504
https://files.catbox.moe/pt8413.png
Its Based64mix-V3

Anonymous

Anonymous Thu 18 May 2023 02:26:58 No.7381521 Report

Quoted By:

>>7381511
>lol, if you don't wanna do it that's fine
different anon

Anonymous

View Same Google ImgOps iqdb SauceNAO 00141.png, 3MiB, 1344x2016

Anonymous Thu 18 May 2023 02:28:31 No.7381526 Report

Quoted By: >>7381825

>>7381516
Thx for replying, and please do. It's an amazing mix

Anonymous

View Same Google ImgOps iqdb SauceNAO tmpco42v0rg.png, 1MiB, 1088x1088

Anonymous Thu 18 May 2023 02:34:19 No.7381538 Report

Quoted By: >>7381904

>>7381455
I love the Fishine lora.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00529-3080114209.jpg, 844KiB, 4100x2811

Anonymous Thu 18 May 2023 02:47:33 No.7381547 Report

Quoted By: >>7381551 >>7381917

using highres in the prompt, seems to help...? at least at 640x640 or higher res images. (it seems to do something with the hands?)

Or it just placebo?

https://files.catbox.moe/2o0m97.png

Anonymous

Anonymous Thu 18 May 2023 02:52:48 No.7381551 Report

Quoted By:

>>7381547
gigacebo

Anonymous

Anonymous Thu 18 May 2023 02:53:23 No.7381552 Report

Quoted By:

>>7379505
this is gorgeous

Anonymous

Anonymous Thu 18 May 2023 02:55:07 No.7381554 Report

Quoted By: >>7381559

>>7379505
catbox?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00149-2221037562.png, 2MiB, 1024x1536

Anonymous Thu 18 May 2023 02:59:47 No.7381559 Report

Quoted By: >>7381560

>>7381554
https://files.catbox.moe/7kcifj.png

Anonymous

Anonymous Thu 18 May 2023 03:02:12 No.7381560 Report

Quoted By: >>7381570

>>7381559
how are you getting it so stylized? im kinda a new fag so im just tryna learn. best im doing rn is like 2.5d stuff but this is lovely

Anonymous

View Same Google ImgOps iqdb SauceNAO 00084-170316253.png, 2MiB, 1024x1536

Anonymous Thu 18 May 2023 03:06:14 No.7381570 Report

Quoted By: >>7381573

>>7381560
you need to get into LoRAs my friend.
There's several on the repost posted in the OP and a bunch more on civit, the one on the catbox you requested is bluethebone and the one from the pic I just posted is the Satoshi Urushihara one from civit.
picrel is the same gen with a different style lora.

Anonymous

Anonymous Thu 18 May 2023 03:09:24 No.7381573 Report

Quoted By: >>7381580

>>7381570
okay ill figure that out i have very few LoRAs, mainly the add detail one and some certain character based ones but i didnt realize they could alter the art style so drastically

Anonymous

View Same Google ImgOps iqdb SauceNAO 00083-170316253.png, 2MiB, 1024x1536

Anonymous Thu 18 May 2023 03:14:16 No.7381580 Report

Quoted By: >>7381585

>>7381573
Style LoRAs are probably the best use for them there is, concept Lora almost always carry SOME style with them that messes with gens, and character Lora most of the times have a hard time getting small details right no matter how good they are.
I've seen very few Character LoRA that manage to get small details right across different positions, outfits and art styles in general.

Anonymous

Anonymous Thu 18 May 2023 03:16:19 No.7381585 Report

Quoted By:

>>7381580
okay later ill try messing with style LoRAs

Anonymous

View Same Google ImgOps iqdb SauceNAO tmp64j0nkpo.png, 1MiB, 1024x1024

Anonymous Thu 18 May 2023 03:43:56 No.7381620 Report

Quoted By: >>7381625

Anonymous

Anonymous Thu 18 May 2023 03:46:48 No.7381625 Report

Quoted By: >>7381627

>>7381620
Teach me that upskirt side shot

Anonymous

Anonymous Thu 18 May 2023 03:48:25 No.7381626 Report

Quoted By: >>7381628

Nice, my 3090 finally arrived. What used to take me 5 minutes in google colab now takes me 24 seconds.

Anonymous

Anonymous Thu 18 May 2023 03:49:25 No.7381627 Report

Quoted By:

>>7381625
https://civitai.com/models/57573/mtu-virus

Anonymous

Anonymous Thu 18 May 2023 03:50:30 No.7381628 Report

Quoted By: >>7381747

>>7381626
>What used to take me 5 minutes in google colab now takes me 24 seconds.
Cumming?

Anonymous

Anonymous Thu 18 May 2023 03:54:19 No.7381633 Report

Quoted By: >>7381642

I wanna experiment with making style LoRAs based on anime screencaps. Do you guys have any suggestions?

Anonymous

Anonymous Thu 18 May 2023 04:00:54 No.7381642 Report

Quoted By: >>7381698

>>7381633
justdoit.gif
scoop up a bunch of style consistent pics, dump them in a folder, and run it through the trainer to see what comes out
then do it again

Anonymous

Anonymous Thu 18 May 2023 04:33:25 No.7381674 Report

Quoted By:

>>7381378
Is the cum part of the gen or was it painted in? If its not something stitched together from multiple images could you share the catbox?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00021-289160288.png, 2MiB, 1280x1536

Anonymous Thu 18 May 2023 04:44:23 No.7381682 Report

Quoted By: >>7381719

Anonymous

Anonymous Thu 18 May 2023 05:14:07 No.7381698 Report

Quoted By: >>7381699

>>7381642
https://github.com/mit-han-lab/fastcomposer

Anonymous

Anonymous Thu 18 May 2023 05:16:16 No.7381699 Report

Quoted By:

>>7381698
didn't mean to (You) anyone, I need sleep

Anonymous

Anonymous Thu 18 May 2023 05:18:57 No.7381700 Report

Quoted By: >>7381725 >>7381796 >>7381824

Which version of SD are people now using? 1.5 or 2.0? I haven't been up to date.

Anonymous

View Same Google ImgOps iqdb SauceNAO tmpspwcm0qv.png, 2MiB, 1536x1024

Anonymous Thu 18 May 2023 05:33:15 No.7381705 Report

Quoted By:

Anonymous

Anonymous Thu 18 May 2023 05:58:52 No.7381719 Report

Quoted By:

>>7381682
catbox?

Anonymous

View Same Google ImgOps iqdb SauceNAO tmpqb7r53hg.png, 3MiB, 2400x1600

Anonymous Thu 18 May 2023 06:00:29 No.7381721 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO 00791-1280759458.png, 2MiB, 1536x1024

Anonymous Thu 18 May 2023 06:02:17 No.7381725 Report

Quoted By:

>>7381700
Don't use SD 2.0. Everything NSFW was removed from its dataset, and LoRAs trained on it can't be used with the webui.

Anonymous

Anonymous Thu 18 May 2023 06:06:45 No.7381730 Report

Quoted By: >>7381742

Is there a reason to make a new line in the prompt window? Such as:
1girl, big honking tooters, blushing
anime screencap, lockerroom background

Anonymous

View Same Google ImgOps iqdb SauceNAO 1678449748965.jpg, 105KiB, 1024x993

Anonymous Thu 18 May 2023 06:08:02 No.7381732 Report

Quoted By: >>7382058

>>7381476
Can you share the cutesexyrobot one as well?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00777-1593686143.png, 2MiB, 1024x1536

Anonymous Thu 18 May 2023 06:20:25 No.7381742 Report

Quoted By: >>7381745

>>7381730
No, it just makes it easier for you to read.

Anonymous

Anonymous Thu 18 May 2023 06:21:51 No.7381745 Report

Quoted By: >>7381749

>>7381742
Huh. But the start and end weight is still applied like usual?

Anonymous

Anonymous Thu 18 May 2023 06:23:26 No.7381747 Report

Quoted By:

>>7381628
That comes later.

I'm doing a fresh install. What are some required extensions?
>regional prompter
>wd 1.4 tagger

Wish I kept a list of extensions I had when I was using colab

Anonymous

View Same Google ImgOps iqdb SauceNAO 01839-1750276089.png, 2MiB, 1536x1024

Anonymous Thu 18 May 2023 06:25:24 No.7381749 Report

Quoted By:

>>7381745
Yes. Only if you insert BREAK will the next token be weighted as if it were at the beginning.

Anonymous

Anonymous Thu 18 May 2023 06:42:59 No.7381766 Report

Quoted By: >>7381813

>>7381353
What lora is this? Looks a bit like Omar dogan

Anonymous

View Same Google ImgOps iqdb SauceNAO 02447-3844585308.jpg, 160KiB, 1024x1536

Anonymous Thu 18 May 2023 06:55:58 No.7381771 Report

Quoted By:

Anonymous

Anonymous Thu 18 May 2023 07:13:53 No.7381783 Report

Quoted By:

>>7381117
link pls

Anonymous

View Same Google ImgOps iqdb SauceNAO 0001.jpg, 560KiB, 1024x1536

Anonymous Thu 18 May 2023 07:29:31 No.7381795 Report

Quoted By: >>7381799

Hi guys,
since there is now a way to store promt-meta in JPG files and use those jpegs in png2image:
is there is some tool/converter available to convert WebUI PNG files to jpgs and keep the meta?
I tried few converters but the meta was stripped :(

Anonymous

Anonymous Thu 18 May 2023 07:30:10 No.7381796 Report

Quoted By: >>7382224

>>7381700
Most people here don't use SD

Anonymous

Anonymous Thu 18 May 2023 07:32:38 No.7381799 Report

Quoted By: >>7381807

>>7381795
I think there's a setting somewhere for saving a jpg version of your image if the png's file size is about a certain threshold
I don't remember if it's only for img2img but it might be what you're looking for, just set the threshold to something like 1 MB

Anonymous

Anonymous Thu 18 May 2023 07:34:39 No.7381800 Report

Quoted By: >>7381852

>git pull
>feels like it runs slower despite updating torch and xformers
wtf. is the other optimization thing worth trying? what was the command? --opt something?

Anonymous

View Same Google ImgOps iqdb SauceNAO 0003.jpg, 507KiB, 1024x1536

Anonymous Thu 18 May 2023 07:38:51 No.7381807 Report

Quoted By: >>7381811

>>7381799
I need a too to convert already created PNGs to JPG format and keep the meta.
Generating new images in JPG format from withing the WebUI is not an issue (just set the output format to jpg there).

Anonymous

Anonymous Thu 18 May 2023 07:40:12 No.7381811 Report

Quoted By: >>7381820

>>7381807
Can't you just recreate the same images then?

Anonymous

Anonymous Thu 18 May 2023 07:42:03 No.7381813 Report

Quoted By:

anyone got suggestions for style loras?

>>7381766
most likely omar-dogan-968it-novae

Anonymous

Anonymous Thu 18 May 2023 07:42:06 No.7381814 Report

Quoted By: >>7381817

Finally figured out the basics of training a LoRA, right now Im generating one for an art style,having it generate a preview every epoch, progress was promising, started to look good then the next one was shit all of a sudden, any tips to avoid this? I did watch a video that said I could select that as the base model and then continue the training afterwards, but I want to know if theres a way to avoid it in the first place

Anonymous

Anonymous Thu 18 May 2023 07:44:49 No.7381817 Report

Quoted By: >>7381833

>>7381814
Enable the option to save every n epochs and then input some reasonable number depending on how many epochs you're doing
You can always just use one of the previous epochs as your lora and delete the ones that come after

Anonymous

View Same Google ImgOps iqdb SauceNAO 00510-2293977764-detailed pupil.png, 2MiB, 1024x1536

Anonymous Thu 18 May 2023 07:45:21 No.7381818 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO 0004.jpg, 395KiB, 1024x1536

Anonymous Thu 18 May 2023 07:48:00 No.7381820 Report

Quoted By: >>7381821

>>7381811
Generating thousands of the same images again?
I'm sorry, are you retarded?

Anonymous

Anonymous Thu 18 May 2023 07:49:27 No.7381821 Report

Quoted By: >>7381827

>>7381820
nta but if you can't figure out a "tool" to edit metadata you might be the retarded one lol

Anonymous

View Same Google ImgOps iqdb SauceNAO 00005-983287450.png, 1MiB, 1024x1536

Anonymous Thu 18 May 2023 07:51:56 No.7381823 Report

Quoted By: >>7381835

Anonymous

Anonymous Thu 18 May 2023 07:58:16 No.7381824 Report

Quoted By:

>>7381700
1.5. they censored 2.0 and it's shit

Anonymous

Anonymous Thu 18 May 2023 08:00:04 No.7381825 Report

Quoted By:

>>7381526
catbox anon?

Anonymous

View Same Google ImgOps iqdb SauceNAO jpg-meta.jpg, 152KiB, 1655x483

Anonymous Thu 18 May 2023 08:02:07 No.7381827 Report

Quoted By: >>7381845

>>7381821
its not your standard exif metadata.They used a a hack to be able to store the promts in jpegs.

Anyway, looks like the ffmpeg can keep the needed block when converting.
Curretly converting 30GB of PNGs to 95% jpegs. Size is ca. 1/5 of the original, while barely loosing quality

Anonymous

Anonymous Thu 18 May 2023 08:14:15 No.7381833 Report

Quoted By:

>>7381817
Some things about the final one are good, eyes are fucked up though, any way to focus on making eyes specifically better like training the model further specifically on images where there's a closeup of the face?

Anonymous

Anonymous Thu 18 May 2023 08:15:27 No.7381835 Report

Quoted By:

>>7381823
double ankle
the chest is facing the wrong way

Anonymous

View Same Google ImgOps iqdb SauceNAO 0005.jpg, 361KiB, 1024x1536

Anonymous Thu 18 May 2023 08:26:57 No.7381845 Report

Quoted By: >>7381871

>>7381827
been reading further on it and looks like it was no hack. The tool in question has to support exif's "UserComment" correctly, to keep the promt info.

Anonymous

Anonymous Thu 18 May 2023 08:32:59 No.7381852 Report

Quoted By: >>7381853

>>7381800
bump. has someone tried a clean install - did it help? how would I even go about uninstalling all of the dependencies?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00372-1103811832-(best quality, (...).png, 2MiB, 1024x1536

Anonymous Thu 18 May 2023 08:34:24 No.7381853 Report

Quoted By: >>7381882

>>7381852
just delete the venv folder, all the deps are in there

Anonymous

Anonymous Thu 18 May 2023 08:36:55 No.7381856 Report

Quoted By: >>7381862 >>7381866 >>7381912

>furry models are several orders of magnitude better at making porn than hentai models
nani the jissai fuck?

Anonymous

View Same Google ImgOps iqdb SauceNAO 2023-05-18 02_48_56-http___local (...).png, 2MiB, 2511x1785

Anonymous Thu 18 May 2023 08:39:19 No.7381861 Report

Quoted By:

kept working on the comfyUI frontend. got an image editor integrated.

Anonymous

Anonymous Thu 18 May 2023 08:41:02 No.7381862 Report

Quoted By:

>>7381856
The furry boorus are impressively autistic with the amount and accuracy of their tags, could be a factor

Anonymous

View Same Google ImgOps iqdb SauceNAO dfws3uy-b04bf508-f8ba-4725-b59e- (...).jpg, 271KiB, 1024x1184

Anonymous Thu 18 May 2023 08:46:01 No.7381865 Report

Quoted By:

Does anyone know how to make stuff like this?
I love these sorta character reference sheets.

Anonymous

Anonymous Thu 18 May 2023 08:48:19 No.7381866 Report

Quoted By: >>7381873

>>7381856
Can you fuck off already?

Anonymous

View Same Google ImgOps iqdb SauceNAO 02503-3399436070.jpg, 167KiB, 1024x1536

Anonymous Thu 18 May 2023 08:50:22 No.7381867 Report

Quoted By: >>7381903

Anonymous

Anonymous Thu 18 May 2023 08:51:02 No.7381868 Report

Quoted By: >>7381869

Is it better to train your loras on models that you plan to actually use them on, or to use the base sd-1.5 one? I've noticed I get better results when Im generating with better shit than sd-1.5 so Im wondering if I should even bother using that as a base to train the lora

Anonymous

Anonymous Thu 18 May 2023 08:52:38 No.7381869 Report

Quoted By:

>>7381868
Train on novelai if you're planning to use the lora with novelai based models

Anonymous

Anonymous Thu 18 May 2023 08:55:41 No.7381871 Report

Quoted By:

>>7381845
holy, catbox?

Anonymous

Anonymous Thu 18 May 2023 08:58:12 No.7381872 Report

Quoted By:

Wouldn't having a futa lora enabled help with inpainting dicks?

Anonymous

Anonymous Thu 18 May 2023 08:59:48 No.7381873 Report

Quoted By:

>>7381866
you first

Anonymous

Anonymous Thu 18 May 2023 09:04:12 No.7381875 Report

Quoted By: >>7381877

I'm cleaning and sorting up my lora folder.
I can't seem to find what 'Hirame-v5' is. And can't seem to find a whipmark lora, needed for thumbnails.

Anyone familiar with these?

Anonymous

Anonymous Thu 18 May 2023 09:06:02 No.7381877 Report

Quoted By: >>7381879

>>7381875
Hirame is the name of an artist so at first glance I'd assume that's what the lora is for
Have you tried checking the tags in the metadata or is it too old to have those?

Anonymous

Anonymous Thu 18 May 2023 09:12:14 No.7381879 Report

Quoted By: >>7381881

>>7381877
Oh shit, didn't know you could see the tags in the safetensor file. Feels like I'm looking at the CLIP tags though?
Either way, most likely that artist, thanks.

Anonymous

Anonymous Thu 18 May 2023 09:15:00 No.7381881 Report

Quoted By: >>7381887

>>7381879
You can use the additional networks tab in the webui to see them in a more digestible format, it also sorts them by most frequently used tags so you (generally) get an idea of which tags will have any effect

Anonymous

View Same Google ImgOps iqdb SauceNAO 00285-2419058608.png, 2MiB, 1024x1536

Anonymous Thu 18 May 2023 09:16:16 No.7381882 Report

Quoted By:

>>7381853

Anonymous

Anonymous Thu 18 May 2023 09:23:52 No.7381887 Report

Quoted By: >>7381889

>>7381881
Damn, out of the 70-ish loras I have, only one had a description, tags or prompts in that additional networks tab..
This must be some well hidden info in all of the lora training guides I guess?

Anonymous

Anonymous Thu 18 May 2023 09:25:26 No.7381889 Report

Quoted By:

>>7381887
It should add them by default nowadays unless someone specifically goes out of their way to remove it, but it's not included in old loras
If a majority of your loras are from the gayshit repository then they were probably just made before the update

Anonymous

Anonymous Thu 18 May 2023 09:38:44 No.7381892 Report

Quoted By:

>trying to inpaint only masked on new verison
>its utter shit now
how does this piece of shit UI continue to get worse with every update

Anonymous

Anonymous Thu 18 May 2023 09:49:17 No.7381903 Report

Quoted By:

>>7381867
old hag gape

Anonymous

Anonymous Thu 18 May 2023 09:50:43 No.7381904 Report

Quoted By:

>>7381538
>Fishine lora
can you share?

Anonymous

Anonymous Thu 18 May 2023 09:54:20 No.7381905 Report

Quoted By:

Some other poorfag using Stable Diffusion via Paperscape? is the thing also stuck updating the webui?

Anonymous

Anonymous Thu 18 May 2023 10:06:59 No.7381912 Report

Quoted By:

>>7381856
They're better than they were.

Anonymous

Anonymous Thu 18 May 2023 10:15:30 No.7381917 Report

Quoted By:

>>7381547
it's not a placebo but it can cook the fuck out of your image

Anonymous

Anonymous Thu 18 May 2023 10:52:37 No.7381935 Report

Quoted By:

G-guys, hello?

>need shortstacks
>can't generate them too good with this model
>look around a bit
>find a fucking lolifying lora
>it actually works to turn stuff into shortstacks

Anonymous

Anonymous Thu 18 May 2023 10:58:23 No.7381942 Report

Quoted By: >>7381946

Does anyone have any idea if there's a bestiality Lora?

Anonymous

Anonymous Thu 18 May 2023 11:06:52 No.7381946 Report

Quoted By: >>7381947

>>7381942
>>>/trash/

Anonymous

Anonymous Thu 18 May 2023 11:07:33 No.7381947 Report

Quoted By: >>7381948

>>7381946
fr wtf was bro thinking asking that shit

Anonymous

Anonymous Thu 18 May 2023 11:08:29 No.7381948 Report

Quoted By: >>7381949

>>7381947
i meant there's an ai thread on trash and since b is another avatarfag general like g and d is just pinups of futas they are the only ones who might now

Anonymous

Anonymous Thu 18 May 2023 11:10:31 No.7381949 Report

Quoted By: >>7381950 >>7381953

>>7381948
still trash anyway, but im a newfag so could you lmk what avatarfag means, this site got a habbit of putting words down then throwing fag onto the end of it "poorfag richfag shitfag aifag etcfag."

Anonymous

Anonymous Thu 18 May 2023 11:11:33 No.7381950 Report

Quoted By: >>7381951

>>7381949
fagfag

Anonymous

Anonymous Thu 18 May 2023 11:12:08 No.7381951 Report

Quoted By:

>>7381950
mf treating me like a retardfag rn

Anonymous

View Same Google ImgOps iqdb SauceNAO Lurk more.jpg, 101KiB, 1000x815

Anonymous Thu 18 May 2023 11:12:42 No.7381953 Report

Quoted By: >>7381956

>>7381949
holy newfag

Anonymous

Anonymous Thu 18 May 2023 11:14:14 No.7381956 Report

Quoted By: >>7381962 >>7381967

>>7381953
stfu oldfag you aint gon come in here chatting that shit im here to learn shit not become a retarded degenerate who's life revolves around this site

Anonymous

Anonymous Thu 18 May 2023 11:21:36 No.7381962 Report

Quoted By:

>>7381956
look at this rebelfag

Anonymous

View Same Google ImgOps iqdb SauceNAO 1.jpg, 526KiB, 2304x1024

Anonymous Thu 18 May 2023 11:25:07 No.7381966 Report

Quoted By: >>7381968 >>7382023

lora mix, pic related, recipe in file name
https://www.mediafire.com/file/3ucy369952w26e7/xmenma_alpha_256_0.5__pokemon_v3_offset_0.3__imoco_stgyki_v1.1-05_0.4__as109_128_dim128_sv_fro_dp0.95_0.4__style-akai_sashimi_0.5_dim32_sv_fro_dp0.95.safetensors/file

Anonymous

View Same Google ImgOps iqdb SauceNAO 00002-409070418.png, 2MiB, 1024x1536

Anonymous Thu 18 May 2023 11:25:40 No.7381967 Report

Quoted By: >>7381969 >>7381974

>>7381956
Go go faggots you can do it fight the others beat those shits!

Anonymous

Anonymous Thu 18 May 2023 11:26:28 No.7381968 Report

Quoted By:

>>7381966
its cute

Anonymous

Anonymous Thu 18 May 2023 11:27:26 No.7381969 Report

Quoted By: >>7381971

>>7381967
looks familiar

Anonymous

Anonymous Thu 18 May 2023 11:29:29 No.7381971 Report

Quoted By:

>>7381969
Blue archive cheerleader outfit lora from civit.ai

Anonymous

Anonymous Thu 18 May 2023 11:31:27 No.7381974 Report

Quoted By:

>>7381967
I was gathering pictures for a general cheerleader lora and like half of the images were either of her or someone in that outfit. Funny thing is, once I got pass her images, the average quality of the images went off a cliff.

Anonymous

Anonymous Thu 18 May 2023 11:41:33 No.7381983 Report

Quoted By:

>>7381376
The detail lora works as a negative, supposedly (I've never used it for that personally)

Anonymous

Anonymous Thu 18 May 2023 11:45:13 No.7381986 Report

Quoted By: >>7381988 >>7381989

Thoughts on easynegative? I heard it can lock the AI into one artstyle and a limited range of compositions compared to negatives like negative_hand, is that true?

Anonymous

Anonymous Thu 18 May 2023 11:47:28 No.7381988 Report

Quoted By:

>>7381986
>Thoughts
There's no reason to use it.

Anonymous

Anonymous Thu 18 May 2023 11:48:46 No.7381989 Report

Quoted By: >>7381990

>>7381986
its shit

Anonymous

Anonymous Thu 18 May 2023 11:52:08 No.7381990 Report

Quoted By:

>>7381989
Yeah that was the impression I got, what of ones like negative_hand?

Anonymous

View Same Google ImgOps iqdb SauceNAO tmpvqf2oxcc.png, 2MiB, 1024x1536

Anonymous Thu 18 May 2023 12:02:43 No.7381997 Report

Quoted By: >>7382011 >>7382019

Results from the LoRA I've been trying to train.

Anonymous

Anonymous Thu 18 May 2023 12:14:36 No.7382011 Report

Quoted By: >>7382019

>>7381997
Isn't a full nelson lora already a thing? Or do you want to improve it?

Anonymous

Anonymous Thu 18 May 2023 12:21:46 No.7382016 Report

Quoted By:

I wish I wasn't so uncreative when it comes to coming up with prompt ideas.

Anonymous

Anonymous Thu 18 May 2023 12:23:41 No.7382019 Report

Quoted By:

>>7381997
>4 hands

>>7382011
The current ones kinda suck

Anonymous

View Same Google ImgOps iqdb SauceNAO 1678597094693372-1738700002-pubi (...).png, 3MiB, 1280x1536

Anonymous Thu 18 May 2023 12:32:27 No.7382023 Report

Quoted By: >>7382027 >>7382034

>>7381966
smells like inpainting. Does it generate workflows?

Anonymous

Anonymous Thu 18 May 2023 12:38:39 No.7382027 Report

Quoted By: >>7382030

>>7382023
you replying to the right person there
can you rephrase

Anonymous

Anonymous Thu 18 May 2023 12:41:26 No.7382030 Report

Quoted By:

>>7382027
lurk more

Anonymous

Anonymous Thu 18 May 2023 12:43:22 No.7382034 Report

Quoted By: >>7382035

>>7382023
FRYINGU

Anonymous

View Same Google ImgOps iqdb SauceNAO 1678597094693383-2802189502-Kaga (...).png, 3MiB, 1280x1536

Anonymous Thu 18 May 2023 12:47:19 No.7382035 Report

Quoted By:

>>7382034
its just the muv luv loras being eternally cursed since CFG was 7.

Anonymous

View Same Google ImgOps iqdb SauceNAO Water.png, 604KiB, 512x768

Anonymous Thu 18 May 2023 12:48:38 No.7382036 Report

Quoted By: >>7382040

Anonymous

View Same Google ImgOps iqdb SauceNAO _petite, medium breasts, solo fo (...).png, 575KiB, 512x768

Anonymous Thu 18 May 2023 12:49:29 No.7382038 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO Water2.png, 579KiB, 512x768

Anonymous Thu 18 May 2023 12:51:57 No.7382039 Report

Quoted By:

Anonymous

Anonymous Thu 18 May 2023 12:52:10 No.7382040 Report

Quoted By: >>7382050 >>7382054

>>7382036
Nice, is that the watercolor lora?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00086-1677755374.png, 526KiB, 640x640

Anonymous Thu 18 May 2023 12:54:22 No.7382043 Report

Quoted By:

How do I resume training with easy scripts? I enabled save state and tried to put lora path next to resume augment in JSON but it just shat me some errors.

Anonymous

View Same Google ImgOps iqdb SauceNAO 1660793983246.png, 586KiB, 512x768

Anonymous Thu 18 May 2023 13:00:10 No.7382046 Report

Quoted By:

Anonymous

Anonymous Thu 18 May 2023 13:02:40 No.7382050 Report

Quoted By: >>7382052 >>7382057 >>7382063

>>7382040
No, just NAI with watercolor and a couple other art style tags.

Anonymous

Anonymous Thu 18 May 2023 13:08:46 No.7382052 Report

Quoted By:

>>7382050
returning to sovl..

Anonymous

Anonymous Thu 18 May 2023 13:10:44 No.7382054 Report

Quoted By:

>>7382040
GRRRRRRR

Anonymous

View Same Google ImgOps iqdb SauceNAO 00016-12602164-1girl, hair ornam (...).png, 3MiB, 1280x1792

Anonymous Thu 18 May 2023 13:16:26 No.7382057 Report

Quoted By:

>>7382050
based

Anonymous

View Same Google ImgOps iqdb SauceNAO 00070-1111817135.png, 2MiB, 1536x1024

Anonymous Thu 18 May 2023 13:17:07 No.7382058 Report

Quoted By: >>7382067 >>7388582

>>7381494
arigatou. It feels good just to have an outlet for my perversion and that my years of experience of browsing boorus can be put to use.

>>7381732
Is it not the same one on gayshit?

Anonymous

Anonymous Thu 18 May 2023 13:21:07 No.7382063 Report

Quoted By:

>>7382050
watercolor is a really nice tag, but it struggles with some things that i guess they don't have images of in watercolor, just gives you something generic instead

Anonymous

View Same Google ImgOps iqdb SauceNAO 00020-2001442595-1girl, hair orn (...).png, 2MiB, 1280x1792

Anonymous Thu 18 May 2023 13:23:37 No.7382065 Report

Quoted By:

prompsmithing was kino and aom/loras killed it

Anonymous

Anonymous Thu 18 May 2023 13:24:10 No.7382066 Report

Quoted By: >>7382068 >>7382069

>>7379197
Definitely worth it, on potato GPUs it's noticeably faster. I still use it on my 3060. I was worried about nondeterministic gens too but it's not like if you reroll it'll make a completely new pic and your waifu is eternally lost to the ether.

>>7379074
NTA, I'm still using vanilla Anything V3. Am I missing out? What's a better upgrade - a different model or loras for stuff I can't get Anything to do (gigantic bolt-on tits, cum facials, deepthroat, dudes not melting their fingers as they grope a girl's breasts, to name a few)

Anonymous

View Same Google ImgOps iqdb SauceNAO 1661908253390.gif, 196KiB, 720x562

Anonymous Thu 18 May 2023 13:25:51 No.7382067 Report

Quoted By: >>7382072

>>7382058
It probably is, its just named different so i asusmed they were different. I've been using your stuff to learn all this. Good shit anon, you have any style presets you use?

Anonymous

Anonymous Thu 18 May 2023 13:27:43 No.7382068 Report

Quoted By:

>>7382066
Newer xformers and sdp are deterministic.
>Anything V3. Am I missing out?
Anon, I fucking beg you... At least use some AOM mix.

Anonymous

Anonymous Thu 18 May 2023 13:28:26 No.7382069 Report

Quoted By: >>7382073

>>7382066
time traveler post

Anonymous

Anonymous Thu 18 May 2023 13:32:18 No.7382072 Report

Quoted By: >>7382091

>>7382067
Nope. I either just load in one of my old prompts or type up something myself.

Anonymous

Anonymous Thu 18 May 2023 13:32:57 No.7382073 Report

Quoted By:

>>7382069
bros out there in actual december
i swear

Anonymous

Anonymous Thu 18 May 2023 13:45:23 No.7382082 Report

Quoted By:

>>7381494
>thanking someone like this for a basic thing everyone itt does
can you suck me off for free too while you're at it?

Anonymous

View Same Google ImgOps iqdb SauceNAO 1674633449306.jpg, 117KiB, 1024x825

Anonymous Thu 18 May 2023 13:54:32 No.7382091 Report

Quoted By: >>7382092

>>7382072
Thanks for the help and props, anon. I'll share some of my stuff later when i get the time.
>(bloodborne) in most prompts
Best souls

Anonymous

Anonymous Thu 18 May 2023 13:54:59 No.7382092 Report

Quoted By:

>>7382091
can you please FUCKING STOP wasting the image limit with those shit images
with sugar on top?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00376-2069790001-nsfw, (pov, 1gi (...).png, 1MiB, 1024x1280

Anonymous Thu 18 May 2023 14:13:16 No.7382102 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO 00050-3133182989.png, 1MiB, 768x1440

Anonymous Thu 18 May 2023 14:18:37 No.7382104 Report

Quoted By: >>7382106

Anonymous

Anonymous Thu 18 May 2023 14:22:14 No.7382106 Report

Quoted By:

>>7382104
>found botw lora earlier today as well as skirt reveal lora

I see you're a man of culture as well.

Anonymous

Anonymous Thu 18 May 2023 14:24:50 No.7382107 Report

Quoted By:

how much additional time do you get with colab pro? honestly considering just doing it for a month or so because my gpu trains slow as shit

Anonymous

View Same Google ImgOps iqdb SauceNAO 22718-457523048-((Ulquiorra)), h (...).png, 829KiB, 1024x1536

Anonymous Thu 18 May 2023 14:28:23 No.7382110 Report

Quoted By: >>7382118

>inb4 where's the hole

Some style loras really improve hands and eyes, I wonder if a model trained on anime frames would be more productive than training on danbooru

Anonymous

View Same Google ImgOps iqdb SauceNAO 00378-3475065715-nsfw, (pov, 1gi (...).png, 2MiB, 1024x1536

Anonymous Thu 18 May 2023 14:29:53 No.7382111 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO 01128-4258797859.png, 2MiB, 1048x1568

Anonymous Thu 18 May 2023 14:30:51 No.7382112 Report

Quoted By: >>7382302

Anonymous

View Same Google ImgOps iqdb SauceNAO 00082-2299634730.png, 2MiB, 1024x1536

Anonymous Thu 18 May 2023 14:31:45 No.7382113 Report

Quoted By: >>7388582

Anonymous

View Same Google ImgOps iqdb SauceNAO 00055-3881854411.png, 4MiB, 2880x1616

Anonymous Thu 18 May 2023 14:37:21 No.7382117 Report

Quoted By: >>7382122

Jesus christ, the ammount of tries I had to role the dice for this one, what's the average amount of tokens you are using for your prompts and negatives?
I feel like lately I'm bloating the process a lot with prompts too big and heavily weighted as well as extensions

Anonymous

Anonymous Thu 18 May 2023 14:39:18 No.7382118 Report

Quoted By: >>7382121 >>7382124

>>7382110
>I wonder if a model trained on anime frames would be more productive than training on danbooru
But few people want to get anime caps.

Anonymous

Anonymous Thu 18 May 2023 14:39:33 No.7382120 Report

Quoted By: >>7382125

so i had to download a bunch of gayshit in an effort to see if i could fix the penis situation since homoerotic art has plenty of uncensored dick,and now my wife thinks im gay.

Anonymous

Anonymous Thu 18 May 2023 14:40:16 No.7382121 Report

Quoted By: >>7382125

>>7382118
im still trying to get ffmpeg to work on python on a frame by frame basis so i don't immiidatly fill out my ssd before i do my other pruning.

Anonymous

Anonymous Thu 18 May 2023 14:40:19 No.7382122 Report

Quoted By:

>>7382117
I dunno, less than 100 for positive and 10 for negative.

Anonymous

Anonymous Thu 18 May 2023 14:41:17 No.7382124 Report

Quoted By: >>7382126

>>7382118
but it's so easy. all you need is ffmpeg.

whatcanimu styles do folks want, anyway?

Anonymous

Anonymous Thu 18 May 2023 14:41:21 No.7382125 Report

Quoted By: >>7382128

>>7382121
I mean gen anime caps.
>>7382120
Homoeroticism works wonder on hands too FYI.

Anonymous

Anonymous Thu 18 May 2023 14:42:01 No.7382126 Report

Quoted By: >>7382136 >>7382241

>>7382124
can you do hidamari sketch?

Anonymous

View Same Google ImgOps iqdb SauceNAO catbox_4igqdv.png, 2MiB, 1024x1536

Anonymous Thu 18 May 2023 14:42:25 No.7382127 Report

Quoted By: >>7382134 >>7382357 >>7382431

Love me this c-string LoRA.

Anonymous

Anonymous Thu 18 May 2023 14:42:35 No.7382128 Report

Quoted By: >>7382129

>>7382125
>Homoeroticism works wonder on hands too FYI.
well thats the least of my problems now.

Anonymous

Anonymous Thu 18 May 2023 14:43:43 No.7382129 Report

Quoted By: >>7382138

>>7382128
You're right on the other part.
I have a male lora that has maybe two penises, one of which is censored, and a sfw style lora and it does dicks and cocksucking better than any female lora I have. Heh.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00449-1833031433.png, 2MiB, 1024x1536

Anonymous Thu 18 May 2023 14:47:09 No.7382134 Report

Quoted By: >>7382135 >>7382189 >>7382431

>>7382127

Anonymous

View Same Google ImgOps iqdb SauceNAO 00252-2737167514.png, 824KiB, 768x1024

Anonymous Thu 18 May 2023 14:48:13 No.7382135 Report

Quoted By: >>7382431

>>7382134

Anonymous

Anonymous Thu 18 May 2023 14:49:10 No.7382136 Report

Quoted By:

>>7382126
>hidamari sketch
I mean it would be done tomorrow or late today at the earliest, are you going yo stick around?

Anonymous

View Same Google ImgOps iqdb SauceNAO 11098-4175799894-((Natsu)), 1gir (...).png, 2MiB, 1024x1536

Anonymous Thu 18 May 2023 14:51:36 No.7382138 Report

Quoted By: >>7382140

>>7382129
>create lora of genderbent character
>add style of theur show
>it now understands how to use the male version of that character in sex scenes

Anonymous

Anonymous Thu 18 May 2023 14:53:01 No.7382140 Report

Quoted By:

>>7382138
I think it's more the yaoi tags since it does 1) the peashooter face and 2) comic style, sometimes; neither of which are present in either lora.
So uh, fags FTW.

Anonymous

Anonymous Thu 18 May 2023 15:01:40 No.7382144 Report

Quoted By: >>7382392

due to request of some guy ages ago i am baking ffx2 rikku, might take a while since i want to try some diff tag setups

Anonymous

View Same Google ImgOps iqdb SauceNAO 00093-2652343473.png, 2MiB, 1024x1536

Anonymous Thu 18 May 2023 15:02:49 No.7382146 Report

Quoted By: >>7382189

Anonymous

View Same Google ImgOps iqdb SauceNAO 00072-1215027820.png, 1MiB, 1024x1536

Anonymous Thu 18 May 2023 15:05:42 No.7382148 Report

Quoted By: >>7382189

the best pose

Anonymous

Anonymous Thu 18 May 2023 15:40:39 No.7382172 Report

Quoted By:

>>7381173
catbox?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00136-466273125.png, 2MiB, 1024x1536

Anonymous Thu 18 May 2023 15:44:16 No.7382173 Report

Quoted By: >>7382180

Anonymous

Anonymous Thu 18 May 2023 15:55:07 No.7382180 Report

Quoted By: >>7382185

>>7382173
catbox?

Anonymous

Anonymous Thu 18 May 2023 15:58:07 No.7382185 Report

Quoted By:

>>7382180
Inpainted the hands, the extras and the face.
https://files.catbox.moe/1wdvuh.png

Anonymous

View Same Google ImgOps iqdb SauceNAO 20230518_114046.jpg, 299KiB, 768x1024

Anonymous Thu 18 May 2023 15:59:31 No.7382186 Report

Quoted By: >>7382189 >>7382191 >>7382192 >>7382193 >>7382204 >>7382390

>>7379025
Yesterday afternoon someone was insisting that is really easy for people to make a lot of money from AI art but I doubt it's that easy. Have any of you made a decent chunk of change from your creations?

Anonymous

Anonymous Thu 18 May 2023 16:01:01 No.7382188 Report

Quoted By:

stop talking about money

Anonymous

Anonymous Thu 18 May 2023 16:01:31 No.7382189 Report

Quoted By: >>7382191 >>7382192 >>7382357

>>7382134
>>7382146
>>7382148
>>7382186
Catbox please?

Anonymous

Anonymous Thu 18 May 2023 16:03:42 No.7382191 Report

Quoted By:

>>7382189
Stole >>7382186 from here:

https://twitter.com/boobamaker/status/1658909154604457988?t=_lZxHtDnFNQ8Ew31wTfPzw&s=19

Anonymous

Anonymous Thu 18 May 2023 16:04:33 No.7382192 Report

Quoted By: >>7382194

>>7382186
Is it time for another federal survey?

>>7382189
Here. Inpainted the face, used img2img to upscale
https://files.catbox.moe/8rx1gv.png

Anonymous

Anonymous Thu 18 May 2023 16:09:36 No.7382193 Report

Quoted By: >>7382194

>>7382186
kys

Anonymous

View Same Google ImgOps iqdb SauceNAO images_12.jpg, 114KiB, 495x619

Anonymous Thu 18 May 2023 16:10:42 No.7382194 Report

Quoted By: >>7382200

>>7382193
>>7382192
Gonna show him these responses the next time that idiot wants to claim you can get rich off of this lol

Anonymous

Anonymous Thu 18 May 2023 16:14:07 No.7382200 Report

Quoted By:

>>7382194
You couldn't glow more if you had a flashlight up your ass

Anonymous

Anonymous Thu 18 May 2023 16:16:59 No.7382204 Report

Quoted By: >>7382212 >>7382523

>>7382186
making about 7000 brazilian real a month by feeding koikatsu images through img2img and selling the result as custom portraits
I think they're using them for D&D or something

Anonymous

Anonymous Thu 18 May 2023 16:23:56 No.7382212 Report

Quoted By:

>>7382204
basado

Anonymous

Anonymous Thu 18 May 2023 16:32:10 No.7382224 Report

Quoted By:

>>7381796
What do you use then if you're not using SD?

Anonymous

Anonymous Thu 18 May 2023 16:37:21 No.7382229 Report

Quoted By:

X/Y/Z plot will create 200 images on 2 4x25 grids. (Total steps to process: 4000)

35min.

Now I wait.

Anonymous

View Same Google ImgOps iqdb SauceNAO catbox_nx7vyl.png, 1MiB, 768x1152

Anonymous Thu 18 May 2023 16:39:42 No.7382232 Report

Quoted By: >>7382255 >>7382291

The CIA are monitoring this thread and keeping tags on all of us. Why? Because the CIA, the FBI and the American Government were the ones to both fund and leak NAI.
But why you ask?
It's simple. The americans want their own crowdsourced AI to combat against Chinese AI and the future of cyberwarfare. But of course they couldn't have been direct and just say it. They needed to do this roundabout way of doing things in order to protect their ego. You see the american government simply ca

Anonymous

View Same Google ImgOps iqdb SauceNAO xyz_grid-0082-976556300-highres, (...).jpg, 789KiB, 3456x1625

Anonymous Thu 18 May 2023 16:43:47 No.7382235 Report

Quoted By: >>7382237 >>7382238 >>7382276 >>7382305

I made a lego man (LM) lora.
I followed the guide on civitai: "Make your own Loras, easy and free".

293 images, 1 rep, 13 epochs.
No activation tag.
unet = 2e-4.

Other settings were the same as the guide.

And I'm getting this purple spot between her hair and breast.
Does anyone know what it is, and how can i fix it?

Anonymous

Anonymous Thu 18 May 2023 16:44:45 No.7382237 Report

Quoted By: >>7382247

>>7382235
hey now, they've gotta look like unfinished sketches for authenticity

Anonymous

Anonymous Thu 18 May 2023 16:45:41 No.7382238 Report

Quoted By: >>7382247

>>7382235
what the fuck are you doing

Anonymous

View Same Google ImgOps iqdb SauceNAO 00000-2826286136.png, 1MiB, 768x1152

Anonymous Thu 18 May 2023 16:45:50 No.7382239 Report

Quoted By:

It's a beautiful day

yay

Anonymous

Anonymous Thu 18 May 2023 16:48:43 No.7382241 Report

Quoted By:

>>7382126
can be done natively, img2img, double or triple the horizontal resolution and check "just resize"

Anonymous

View Same Google ImgOps iqdb SauceNAO 325682937544839518.png, 1MiB, 864x1304

Anonymous Thu 18 May 2023 16:53:06 No.7382244 Report

Quoted By:

Anonymous

Anonymous Thu 18 May 2023 16:56:33 No.7382247 Report

Quoted By: >>7382278

>>7382237
i filtered every non-colorized picture.
Maybe was the different render/paint styles that caused the purple spots?

>>7382238
A lora. If you know how I could fix the problem I stated, it would be of great help.

The X asis is the epoch, the Y is the lora value in the prompt.

Anonymous

Anonymous Thu 18 May 2023 17:01:37 No.7382249 Report

Quoted By: >>7382251

haven't been here in like a month, what's the hottest, biggest and bestest model right now?

Anonymous

Anonymous Thu 18 May 2023 17:02:22 No.7382251 Report

Quoted By: >>7382262

>>7382249
AOM2

Anonymous

View Same Google ImgOps iqdb SauceNAO 00095-612244517.png, 2MiB, 1280x1536

Anonymous Thu 18 May 2023 17:02:24 No.7382252 Report

Quoted By:

Anonymous

Anonymous Thu 18 May 2023 17:04:06 No.7382255 Report

Quoted By:

>>7382232
Hey, this is a crock of shit!
<Please remove this part of the message as part of OPSEC Protocol TS.>

Anonymous

View Same Google ImgOps iqdb SauceNAO file.png, 1MiB, 813x986

Anonymous Thu 18 May 2023 17:07:30 No.7382258 Report

Quoted By: >>7382275

>>7379579
Hell, which fanboxes have lora resources and other useful things like that?

Anonymous

Anonymous Thu 18 May 2023 17:09:31 No.7382262 Report

Quoted By: >>7382263 >>7382268

>>7382251
>AOM2
Which is Aom2? I don't see it in any of the links. Is it a model recipe?

Anonymous

Anonymous Thu 18 May 2023 17:09:55 No.7382263 Report

Quoted By:

>>7382262
Yes

Anonymous

Anonymous Thu 18 May 2023 17:11:54 No.7382267 Report

Quoted By:

>>7379707
Amazing 2D hentai image

...

...

catbox.

Anonymous

Anonymous Thu 18 May 2023 17:12:08 No.7382268 Report

Quoted By:

>>7382262
Mhm, gotta cook it up yourself. It's sekrit sauce.

Anonymous

View Same Google ImgOps iqdb SauceNAO image_2023-05-18_131659287.png, 569KiB, 768x768

Anonymous Thu 18 May 2023 17:17:07 No.7382274 Report

Quoted By:

Anonymous

Anonymous Thu 18 May 2023 17:17:22 No.7382275 Report

Quoted By:

>>7382258
thats a nice bun. catbox?

Anonymous

Anonymous Thu 18 May 2023 17:18:22 No.7382276 Report

Quoted By: >>7382277 >>7382286

>>7382235
>assuming batch size 1, 3809it
is this bait or are people here this retarded now? i swear everyday we regress further and further; it almost feels like the weeks after loras dropped, except without the useful anons experimenting and just retards peddling stupid shit

Anonymous

Anonymous Thu 18 May 2023 17:19:04 No.7382277 Report

Quoted By: >>7382282

>>7382276
good morning sir

Anonymous

Anonymous Thu 18 May 2023 17:20:22 No.7382278 Report

Quoted By: >>7382290

>>7382247
Most of his colorized shit isn't even by him and all of it sucks, train on his sketches

Anonymous

View Same Google ImgOps iqdb SauceNAO 1666804013511128-1817444323-mast (...).png, 1MiB, 1024x1536

Anonymous Thu 18 May 2023 17:21:35 No.7382280 Report

Quoted By: >>7382285 >>7382306 >>7382311

Updooted my mefa with more comprehensibke names. nit using preview images as anons duggested against it for copyright.

https://mega.nz/folder/xDkWWRAZ#jrvHsIeim3Q1HyDUEHLjbw

Anonymous

Anonymous Thu 18 May 2023 17:23:39 No.7382282 Report

Quoted By: >>7382287

>>7382277
i don't get the implication

Anonymous

View Same Google ImgOps iqdb SauceNAO 00123-3036455823.png, 2MiB, 1280x1536

Anonymous Thu 18 May 2023 17:25:18 No.7382283 Report

Quoted By: >>7382299

Anonymous

Anonymous Thu 18 May 2023 17:26:56 No.7382285 Report

Quoted By:

>>7382280
based

Anonymous

Anonymous Thu 18 May 2023 17:26:59 No.7382286 Report

Quoted By: >>7382292

>>7382276
Batch size was 2. I asked for help because I'm still learning.

But I see you are just an autistic screeching retard.

Anonymous

Anonymous Thu 18 May 2023 17:27:44 No.7382287 Report

Quoted By:

>>7382282
Well, you just woke up. I want you to have a good day.

Anonymous

Anonymous Thu 18 May 2023 17:29:02 No.7382290 Report

Quoted By:

>>7382278
Yeah, I know. The colorized pictures I got were ok for the most part.

I was trying to get the line/form style in general. I think it worked, but of course it can be improved.

I will try the sketches too.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00284-3116436522.png, 2MiB, 1024x1536

Anonymous Thu 18 May 2023 17:29:08 No.7382291 Report

Quoted By:

>>7382232

Anonymous

Anonymous Thu 18 May 2023 17:29:55 No.7382292 Report

Quoted By:

>>7382286
>But I see you are just an autistic screeching retard.
lmao fuck off and go back for to civitai then

Anonymous

Anonymous Thu 18 May 2023 17:31:06 No.7382296 Report

Quoted By: >>7382297 >>7382300 >>7382316

I'm still using based65

Anonymous

Anonymous Thu 18 May 2023 17:31:55 No.7382297 Report

Quoted By:

>>7382296
chad

Anonymous

Anonymous Thu 18 May 2023 17:33:12 No.7382299 Report

Quoted By:

>>7382283
>that smugness

Anonymous

Anonymous Thu 18 May 2023 17:33:55 No.7382300 Report

Quoted By:

>>7382296
Are you implying that you should be using something else to fit the norm? Completely serious here since i haven't kept up with models at all.

Anonymous

Anonymous Thu 18 May 2023 17:37:22 No.7382302 Report

Quoted By:

>>7382112
catbox?

Anonymous

View Same Google ImgOps iqdb SauceNAO 1664907906839.png, 603KiB, 512x768

Anonymous Thu 18 May 2023 17:38:16 No.7382303 Report

Quoted By:

Anonymous

Anonymous Thu 18 May 2023 17:38:30 No.7382305 Report

Quoted By: >>7382323

>>7382235
Bare minimum when asking for help is to post all your settings from additional networks tab or wherever. Nobody's going to read some civitai guide just to get all the pieces of the puzzle.

Anonymous

Anonymous Thu 18 May 2023 17:38:31 No.7382306 Report

Quoted By:

>>7382280
Fuck me I hate this keyboard

Anonymous

View Same Google ImgOps iqdb SauceNAO 1662557650876.png, 537KiB, 512x768

Anonymous Thu 18 May 2023 17:38:46 No.7382307 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO 1666515499517.png, 566KiB, 512x768

Anonymous Thu 18 May 2023 17:39:32 No.7382309 Report

Quoted By:

Anonymous

Anonymous Thu 18 May 2023 17:42:12 No.7382311 Report

Quoted By:

>>7382280
gj

Anonymous

Anonymous Thu 18 May 2023 17:45:49 No.7382316 Report

Quoted By:

>>7382296
proud of you king

Anonymous

Anonymous Thu 18 May 2023 17:49:47 No.7382319 Report

Quoted By: >>7382322 >>7382341

https://github.com/ljleb/sd-webui-neutral-prompt
what does this mean for my cock

Anonymous

Anonymous Thu 18 May 2023 17:56:22 No.7382322 Report

Quoted By: >>7382324

>>7382319
I just got it and I don't really understand what it's supposed to do.

Anonymous

Anonymous Thu 18 May 2023 17:56:32 No.7382323 Report

Quoted By: >>7382349

>>7382305
No problem. For what i saw, i though there was some kind of "standard" in lora/loncon creation.

Specifically Tried Loncon/Lycoris Structure.
Here's the pastebin with all info from additional networks tab: https://pastebin.com/GN6CrMMU

Settings used in the colab:
293 images, 1 rep, 13 epochs.
No activation tag.
unet = 2e-4.
text encoder = 1e-4
shuffle_tags = activated
scheduler = cosine with restarts
lr_scheduler_number = 3
warmup ratio = 3
network_dim = 16
network_alpha = 8
conv_dim = 8
conv_alpha = 1

Anonymous

Anonymous Thu 18 May 2023 17:58:11 No.7382324 Report

Quoted By: >>7382326

>>7382322
Wait shit, I kinda get it???
That's dope.

Anonymous

Anonymous Thu 18 May 2023 17:59:14 No.7382326 Report

Quoted By: >>7382335

>>7382324
No I don't, never mind.

Anonymous

Anonymous Thu 18 May 2023 18:07:43 No.7382335 Report

Quoted By: >>7382338

>>7382326
this is the real shit part of that extension. it fixes the negatives to actually work properly through the power of maths or something

https://perp-neg.github.io/

Anonymous

Anonymous Thu 18 May 2023 18:11:12 No.7382338 Report

Quoted By: >>7382341 >>7382342

>>7382335
Oh. I thought it was supposed to be CFG. So how do you actually use it?

Anonymous

Anonymous Thu 18 May 2023 18:12:41 No.7382341 Report

Quoted By: >>7382345

>>7382319
It's from that new paper that's been shared around the past couple days. Looks like it adds an additional one too that I can't say I'm familiar with.
>>7382338
You install the extension?

Anonymous

Anonymous Thu 18 May 2023 18:12:44 No.7382342 Report

Quoted By: >>7382343

>>7382338
it doesn't look like it has an off button, it just changes the cfg algorithm. and honestly shit looks better already

Anonymous

Anonymous Thu 18 May 2023 18:13:53 No.7382343 Report

Quoted By: >>7382347

>>7382342
You people are braindead. Yes it does. https://github.com/ljleb/sd-webui-neutral-prompt/blob/f58f13e8ce44d7d7f86457f848ff4960105ffc73/scripts/main.py#L101

Anonymous

Anonymous Thu 18 May 2023 18:16:16 No.7382345 Report

Quoted By: >>7382348

>>7382341
>You install the extension?
And then? It does shit if you don't tweak the controls.

Anonymous

Anonymous Thu 18 May 2023 18:17:36 No.7382347 Report

Quoted By:

>>7382343
yeah i tabbed back and realized im retarded

Anonymous

Anonymous Thu 18 May 2023 18:17:48 No.7382348 Report

Quoted By: >>7382350

>>7382345
Read the paper(s) or experiment with the controls yourself, silly anon. Maybe wait for this PR to be addressed because I think the maths might actually be wrong in it's current state. https://github.com/ljleb/sd-webui-neutral-prompt/pull/2

Anonymous

Anonymous Thu 18 May 2023 18:19:09 No.7382349 Report

Quoted By:

>>7382323
Settings look pretty normal other than text encoder being high at 1e4, at least with normal loras it's usually in e5 range. No clue about lycoris but for new people I'd recommend starting with a normal ass lora with 1:1 network dim and alpha at high values.

Definitely up your training resolution from 512 to 768 if you can. Should be possible even for poor people with some vram saving options like gradient_checkpointing, cache_latents_to_disk etc.

Anonymous

Anonymous Thu 18 May 2023 18:19:57 No.7382350 Report

Quoted By: >>7382358 >>7382359

>>7382348
its not ridiculous to expect at least a basic description of what each feature in an extension does

Anonymous

View Same Google ImgOps iqdb SauceNAO 00379-1470480332-best quality, p (...).png, 2MiB, 1024x1536

Anonymous Thu 18 May 2023 18:22:13 No.7382355 Report

Quoted By:

hmhmhm

Anonymous

View Same Google ImgOps iqdb SauceNAO 00304-2393675149.jpg, 3MiB, 3048x1488

Anonymous Thu 18 May 2023 18:22:44 No.7382356 Report

Quoted By: >>7382364 >>7383460

Anonymous

Anonymous Thu 18 May 2023 18:22:58 No.7382357 Report

Quoted By:

>>7382189
The prompts are very similar to >>7382127

The image you want is a combination of these two, inpainted to make them blend together.
https://files.catbox.moe/lkgct4.png
https://files.catbox.moe/sntct8.png

The initial body was generated with this Sakura LoRA:
https://civitai.com/models/67082/haruno-sakura-naruto-character-lora

And then I inpainted the face with this Sakura LoRA (which is actually trained on her forehead mark):
https://civitai.com/models/6905/haruno-sakura-naruto-lora

Here's the c-string LoRA while I'm at it:
https://civitai.com/models/9972/c-string

All of this was necessary because I still haven't found a way to get skirts that only go on the sides but leave the front uncovered. I've tried "showgirl skirt", "half skirt", and "detached skirt", none of it works. For the final picture, I didn't even want that red belt attaching the two sides together, but the AI wasn't having it.

Anonymous

Anonymous Thu 18 May 2023 18:23:22 No.7382358 Report

Quoted By: >>7382360 >>7382362 >>7382371

>>7382350
CFG rescale should be 0.75 according to the paper. I haven't read the other paper so I can't answer for the rest. Usually these don't have a simple explantation because they're very math heavy and are meant to act as replacements and not to be exposed to the end-user.

Anonymous

Anonymous Thu 18 May 2023 18:23:26 No.7382359 Report

Quoted By:

>>7382350
Yes, especially when someone advertises it under "nut destroy" pretenses. Oh well.

Anonymous

Anonymous Thu 18 May 2023 18:25:54 No.7382360 Report

Quoted By:

>>7382358
I guess it does work to make CFG not change the composition, but it just fucks the images BAR, so not for me.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00380-2913279250-best quality, p (...).png, 2MiB, 1024x1536

Anonymous Thu 18 May 2023 18:26:38 No.7382361 Report

Quoted By:

Anonymous

Anonymous Thu 18 May 2023 18:26:49 No.7382362 Report

Quoted By: >>7382371

>>7382358
All it seemed to do for me is suck all color from my gens so that's cool. Going to wait and see if people figure this out

Anonymous

View Same Google ImgOps iqdb SauceNAO 00197-2954268930.png, 2MiB, 1024x1536

Anonymous Thu 18 May 2023 18:27:48 No.7382364 Report

Quoted By: >>7382365 >>7388582

>>7382356
impressive girth.

Anonymous

Anonymous Thu 18 May 2023 18:28:14 No.7382365 Report

Quoted By:

>>7382364
i like this one

Anonymous

Anonymous Thu 18 May 2023 18:31:28 No.7382368 Report

Quoted By: >>7382381

is there an specific prompt for thin legs?

Anonymous

Anonymous Thu 18 May 2023 18:32:35 No.7382371 Report

Quoted By:

>>7382358
>>7382362
I took a cursory look at the other paper and the extension's math is indeed fucked up. Wait until https://github.com/ljleb/sd-webui-neutral-prompt/pull/2 is dealt with.

Anonymous

View Same Google ImgOps iqdb SauceNAO 1684324769306372-794254013-full (...).png, 1MiB, 768x1152

Anonymous Thu 18 May 2023 18:40:02 No.7382380 Report

Quoted By:

yeap it's time to rebake at lower lr and specify colors in all the tags, it didn't get the boots more than 20% of the time until the cook zone

Anonymous

View Same Google ImgOps iqdb SauceNAO catbox_460olx.png, 2MiB, 1024x1536

Anonymous Thu 18 May 2023 18:40:05 No.7382381 Report

Quoted By: >>7382382 >>7382384

>>7382368
Have you tried "thin", "skinny", "legs", "long legs", "skinny legs", or "thin legs"?

Come on anon, this took me no time to make. It's not rocket science.

Anonymous

Anonymous Thu 18 May 2023 18:40:41 No.7382382 Report

Quoted By: >>7382389

>>7382381
you got a point but bruhhhhhh that image is cooked af

Anonymous

Anonymous Thu 18 May 2023 18:41:45 No.7382384 Report

Quoted By: >>7382389

>>7382381
oh it kinda works
I was only using thin and thin legs
thx

Anonymous

View Same Google ImgOps iqdb SauceNAO 00383-3509799362-best quality, p (...).png, 3MiB, 1536x1536

Anonymous Thu 18 May 2023 18:45:00 No.7382388 Report

Quoted By:

hmmm, I resumed training for a bunch more epochs but it didn't change at all, how do I fry more style

Anonymous

View Same Google ImgOps iqdb SauceNAO 00455-4033939789.png, 2MiB, 1024x1536

Anonymous Thu 18 May 2023 18:45:25 No.7382389 Report

Quoted By: >>7382394

>>7382382
>>7382384

Uncooked it. Whenever I do a person full body head to toes I always have to inpaint, don't know why.

Anonymous

View Same Google ImgOps iqdb SauceNAO file.png, 6KiB, 1181x73

Anonymous Thu 18 May 2023 18:46:34 No.7382390 Report

Quoted By: >>7382530

>>7382186
denied like 4 requests before closing down and got 11 fanbox followers so you can do the math. And im not doing anything popular either, the dudes making mainstream shit were making insane easy money, you can go through pixic profiles and see how many requests they had or look up fat patreons

Anonymous

Anonymous Thu 18 May 2023 18:47:16 No.7382392 Report

Quoted By: >>7382397

>>7382144
Cool idea anon, I'm surprised I haven't seen any Rikku gens yet

Anonymous

Anonymous Thu 18 May 2023 18:48:00 No.7382394 Report

Quoted By: >>7382401 >>7382404 >>7382431

>>7382389
what the fuck are you doing

>(best quality, ultra-detailed, best illustration, best shadow, masterpiece, high res), detailed background
>verybadimagenegative_v1.3, EasyNegative, ng_deepnegative_v1_75t, (painting by bad-artist-anime:0.9), (painting by bad-artist:0.9)
>CFG scale: 12

brother

Anonymous

Anonymous Thu 18 May 2023 18:48:45 No.7382396 Report

Quoted By: >>7382404 >>7382431

oh and
>clip 1

Anonymous

View Same Google ImgOps iqdb SauceNAO 1684324769306377-3632190583-full (...).png, 1MiB, 768x1152

Anonymous Thu 18 May 2023 18:49:23 No.7382397 Report

Quoted By:

>>7382392
soon tm

Anonymous

Anonymous Thu 18 May 2023 18:49:53 No.7382401 Report

Quoted By: >>7382404 >>7382431

>>7382394
>Steps: 36
>Sampler: DPM++ SDE Karras
lol

Anonymous

Anonymous Thu 18 May 2023 18:51:46 No.7382404 Report

Quoted By: >>7382431

>>7382394
>>7382396
>>7382401
GIGABASED
holy

Anonymous

View Same Google ImgOps iqdb SauceNAO 1684324769306379-3673601893-1gir (...).png, 2MiB, 1024x1536

Anonymous Thu 18 May 2023 19:00:11 No.7382418 Report

Quoted By: >>7382426 >>7382431 >>7382740

the tag is "halterneck" btw, it's a little unreliable

don't want to beat a dead horse but you really don't need all those embeds obliterating the lines on your image, especially if you're going to inpaint anyway

Anonymous

Anonymous Thu 18 May 2023 19:05:44 No.7382426 Report

Quoted By: >>7382431

>>7382418
nta but I also have like 4-5 embeds
I've done enough tests and I like having these

Anonymous

View Same Google ImgOps iqdb SauceNAO 00176-3526471336.png, 1004KiB, 768x1024

Anonymous Thu 18 May 2023 19:08:13 No.7382431 Report

Quoted By: >>7382434 >>7382437 >>7382439 >>7382448

>>7382394
>>7382396
>>7382401
>>7382404
>>7382418
>>7382426

The CFG scale and clip 1 are mistakes. I was messing around with some other models earlier and forgot to switch back. And some prompts just get left in there for a while like "censored:1.5".

I'll be honest, I just throw shit in and then forget about it. All my prompts end up being super wordy and it works out pretty well imo.

I made >>7382127 >>7382134 >>7382135
The Kyoko one didn't even need inpainting, that's straight txt2img.

I think I do an alright job.

What would people recommend anyway? Just one or two negative embeds?

I also don't really know what's wrong with my sampler or number of steps. I mostly just mess around until I get something I like and steal shit from other people's metadata, I never read any guides on what best practices are.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00062-316774398-nsfw, scan, (tra (...).png, 2MiB, 1024x1536

Anonymous Thu 18 May 2023 19:08:26 No.7382433 Report

Quoted By: >>7382482

Anonymous

Anonymous Thu 18 May 2023 19:09:16 No.7382434 Report

Quoted By: >>7382458

>>7382431
look at the edges of the lines, they are fuzzy, are you blind or someth

you dont need any negative embeds imo, though i havent heard a lot of complaints about bad hands

Anonymous

Anonymous Thu 18 May 2023 19:10:51 No.7382437 Report

Quoted By: >>7382458

>>7382431
>I also don't really know what's wrong with my sampler or number of steps.
SDE is just slower than most, karras gets the job done faster with the same amount of steps

for the rest it's just taste, try to change/rework some prompts and do tests or whatever

I reworked ALL my prompts like a month ago and went from shitty NAI-looking results to actual good results
but I still use a lot of prompts and embeds though

Anonymous

Anonymous Thu 18 May 2023 19:13:15 No.7382439 Report

Quoted By: >>7382458

>>7382431
>no negative embeds
>aom negative
>aom mix
>cfg under 7
>no positive placebos
>15 steps for sde karras
something like that

Anonymous

Anonymous Thu 18 May 2023 19:22:00 No.7382448 Report

Quoted By: >>7382458 >>7382475

>>7382431
>Just one or two negative embeds?
0.
>steps
karras needs less, like 15
>cfg
6-8. Even in the olden days 11 was peak, its not exactly some strict rule but 12's a bit much

Anonymous

Anonymous Thu 18 May 2023 19:30:26 No.7382458 Report

Quoted By: >>7382462 >>7382464 >>7382478

>>7382434
>>7382437
>>7382439
>>7382448
I'll take it all under advisement. It's a bit moot since I inpaint to fix most stuff anyway, but it would be nice to have fewer inpainting edges.

I'll need to do more tweaking though, I'm seeing... worse results.

Tried with and without the "positive placebos", not seeing a difference, so I guess they are just placebo.

Anonymous

Anonymous Thu 18 May 2023 19:32:22 No.7382462 Report

Quoted By:

>>7382458
Image max was hit, so here's a URL instead

https://files.catbox.moe/ilfxhq.png

Anonymous

Anonymous Thu 18 May 2023 19:33:23 No.7382464 Report

Quoted By: >>7382467

>>7382458
Image max was hit, so here's a URL instead
https://files.catbox.moe/b3l7kb.png

Anonymous

Anonymous Thu 18 May 2023 19:33:31 No.7382465 Report

Quoted By: >>7382471

so
nothing interesting to make into a theme again huh?

Anonymous

Anonymous Thu 18 May 2023 19:34:32 No.7382467 Report

Quoted By:

>>7382464
wish that was bait

Anonymous

Anonymous Thu 18 May 2023 19:36:04 No.7382468 Report

Quoted By: >>7382469

https://litter.catbox.moe/jc5h17.png

Anonymous

Anonymous Thu 18 May 2023 19:37:02 No.7382469 Report

Quoted By: >>7382476

>>7382468
That's some proper schizoprompting, good mate.

Anonymous

Anonymous Thu 18 May 2023 19:38:33 No.7382471 Report

Quoted By: >>7382472

>>7379239
>>7382465
ahem

Anonymous

Anonymous Thu 18 May 2023 19:39:38 No.7382472 Report

Quoted By:

>>7382471
...
BASED

Anonymous

Anonymous Thu 18 May 2023 19:41:01 No.7382475 Report

Quoted By: >>7382479

>>7382448
I still like a cfg of ~12 personally, especially with euler a

Anonymous

Anonymous Thu 18 May 2023 19:41:14 No.7382476 Report

Quoted By:

>>7382469
I have tried regular tags and expressive sentences. I find the sentences work best but with a high signal:noise ratio. Sometimes it drops beautiful images without any inpainting needing, other times it's a monster hybrid.

Anonymous

Anonymous Thu 18 May 2023 19:42:08 No.7382478 Report

Quoted By: >>7382481

>>7382458
its not exactly placebo, style prompts still matter but you gotta know which do shit and which dont, then use a model thats not overfit

Anonymous

Anonymous Thu 18 May 2023 19:42:39 No.7382479 Report

Quoted By:

>>7382475
I'll only do that in img2img with a very basic prompt like ((best quality, masterpiece, realistic)), (detailed)

Anonymous

Anonymous Thu 18 May 2023 19:43:38 No.7382481 Report

Quoted By: >>7382499

>>7382478
I don't think he got what I meant by positive placebos judging by that last catbox.

Anonymous

Anonymous Thu 18 May 2023 19:43:51 No.7382482 Report

Quoted By:

>>7382433
phew...

Anonymous

Anonymous Thu 18 May 2023 19:45:53 No.7382484 Report

Quoted By: >>7382486 >>7382493 >>7382514

oh fuck i have an idea for thread theme

Anonymous

Anonymous Thu 18 May 2023 19:47:34 No.7382486 Report

Quoted By: >>7382494 >>7382514

>>7382484
You do? Because I was about to bake.

Anonymous

Anonymous Thu 18 May 2023 19:49:44 No.7382489 Report

Quoted By: >>7382495 >>7382497 >>7382499

Placebo implies it has no effect. Those tags you call "placebo" have an effect, whether it's more or less noticeable depends on both the tag and the model. Whether the tag does what you think it does is a whole other matter entirely.

Anonymous

Anonymous Thu 18 May 2023 19:49:47 No.7382490 Report

Quoted By:

Why is every image prompt on CivitAI so schizo?

Anonymous

Anonymous Thu 18 May 2023 19:50:42 No.7382493 Report

Quoted By:

>>7382484
pregnant elves

Anonymous

Anonymous Thu 18 May 2023 19:50:45 No.7382494 Report

Quoted By: >>7382514

>>7382486
yeah 1 sec

Anonymous

Anonymous Thu 18 May 2023 19:50:54 No.7382495 Report

Quoted By:

>>7382489
I can call them useless failures, but placebo rolls better off the tongue.

Anonymous

Anonymous Thu 18 May 2023 19:50:57 No.7382497 Report

Quoted By:

>>7382489
this

Anonymous

Anonymous Thu 18 May 2023 19:51:32 No.7382499 Report

Quoted By: >>7382504

>>7382489
>Whether the tag does what you think it does is a whole other matter entirely.
Thats what we mean by "placebo" here dumbass. "quality" tags not affecting quality that is.

>>7382481
>looked
its over..

Anonymous

Anonymous Thu 18 May 2023 19:52:28 No.7382504 Report

Quoted By: >>7382506 >>7382510

>>7382499
placebo has an actual definition you can't just change it to whatever the fuck you want

Anonymous

Anonymous Thu 18 May 2023 19:53:16 No.7382506 Report

Quoted By: >>7382511

>>7382504
>Something of no intrinsic remedial value that is used to appease or reassure another
based dictionary

Anonymous

Anonymous Thu 18 May 2023 19:53:43 No.7382510 Report

Quoted By: >>7382511

>>7382504
nta but from oxford
>a measure designed merely to calm or please someone.

Anonymous

Anonymous Thu 18 May 2023 19:54:20 No.7382511 Report

Quoted By: >>7382512 >>7382526

>>7382506
>>7382510
but those tags actually do things, they just don't do what they say they do. they are mislabeled or misused, not nonfucntional or fake, which is what placebo means

Anonymous

Anonymous Thu 18 May 2023 19:54:55 No.7382512 Report

Quoted By: >>7382517

>>7382511
>no intristic remedial value

Anonymous

Anonymous Thu 18 May 2023 19:55:30 No.7382514 Report

Quoted By:

>>7382484
>>7382486
>>7382494
https://imgur.com/a/tUA3rAK

Anonymous

Anonymous Thu 18 May 2023 19:56:27 No.7382517 Report

Quoted By: >>7382521

>>7382512
you don't understand what those words mean clearly

Anonymous

Anonymous Thu 18 May 2023 19:57:30 No.7382521 Report

Quoted By:

>>7382517
I'd get you some dictionary definitions but you couldn't understand one already, so...

Anonymous

Anonymous Thu 18 May 2023 19:57:58 No.7382523 Report

Quoted By:

>>7382204
>custom portraits
What kind of market even is that? You mean physical portraits or something? What site do you even go to for that?

Anonymous

Anonymous Thu 18 May 2023 19:58:39 No.7382526 Report

Quoted By: >>7382533

>>7382511
non functional or fake are not synonyms of placebo, anon

Anonymous

Anonymous Thu 18 May 2023 19:59:27 No.7382530 Report

Quoted By: >>7382570

>>7382390
Link your account so I can see what type of shit actually makes money.... Pweeze?

Anonymous

Anonymous Thu 18 May 2023 20:00:13 No.7382533 Report

Quoted By:

>>7382526
the definition is that the ONLY use of them is to calm people or make them think that something is happening in order to make them feel better. that is not what is happening here.

Anonymous

Anonymous Thu 18 May 2023 20:00:43 No.7382538 Report

Quoted By:

it's not a fucking placebo it's just a shit tag

Anonymous

Anonymous Thu 18 May 2023 20:01:22 No.7382543 Report

Quoted By:

BREAD
BREAD
BREAD
>>7382540
>>7382540
>>7382540

Anonymous

Anonymous Thu 18 May 2023 20:11:33 No.7382570 Report

Quoted By: >>7382590

>>7382530
lurk more. its just fans of a single series in my case, you're better off looking up actual popular accounts

Anonymous

Anonymous Thu 18 May 2023 20:23:42 No.7382590 Report

Quoted By: >>7382598

>>7382570
By serious do you mean a certain obscure anime or your own OCs?

Anonymous

Anonymous Thu 18 May 2023 20:27:02 No.7382598 Report

Quoted By:

>>7382590
obscure

Anonymous

Anonymous Thu 18 May 2023 21:38:34 No.7382740 Report

Quoted By:

>>7382418
catbox?

Anonymous

Anonymous Thu 18 May 2023 22:09:41 No.7382786 Report

Quoted By:

>>>/vg/429688447
Artificial Academy 2 General /aa2g/ #1279
Who Framed /aa2g/? Edition

Welcome, this general is for the discussion of ILLUSION's Artificial Academy 2.

COPY ERROR MESSAGES WITH CTRL+C, PASTE THEM WITH CTRL+V INTO GOOGLE TRANSLATE. JUST CLICK THE WINDOW AND PRESS CTRL + C, IT WORKS.

>Downloads:
/aa2g/ Pre-Installed Game, AA2Mini: https://tsukiyo.me/AAA/AA2MiniPPX.xml
AAUnlimited updates: https://github.com/aa2g/AA2Unlimited/releases
Anon's Modded Pre-Install: https://pastebin.com/42JS3q6E

>Information:
AA2Mini Install Guide:
https://docs.google.com/document/d/e/2PACX-1vS8Ap6CrmSNXRsKG9jsIMqHYuHM3Cfs5qE5nX6iIgfzLlcWnmiwzmOrp27ytEMX03lFNRR7U5UXJalA/pub
General FAQ:
https://web.archive.org/web/20200216045726/https://pastebin.com/bhrA6iGx
AAU Guide and Resources (Modules, Tans, Props, Poses, and More):
https://docs.google.com/spreadsheets/d/17qb1X0oOdMKU4OIDp8AfFdLtl5y_4jeOOQfPQ2F-PKQ/edit#gid=0

>Character Cards [Database], now with a list of every NonOC in the megas:
https://docs.google.com/spreadsheets/d/1niC6g-Xd2a2yaY98NBFdAXnURi4ly2-lKty69rkQbJ0/edit#gid=2085826690
https://db.bepis.moe/aa2/

>Mods & More:
Mods for AAU/AA2Mini (ppx format, the mediafire has everything):
https://www.mediafire.com/folder/vwrmdohus4vhh/Mods
/aa2g/ Modding Reference Guide (Slot lists for Hair/Clothes/Faces, List Guides, and More):
https://docs.google.com/spreadsheets/d/1gwmoVpKuSuF0PtEPLEB17eK_dexPaKU106ShZEpBLhg/edit#gid=1751233129
Booru: https://aau.booru.org

>HELP! I have a Nvidia card and my game crashes on startup!
Try the dgVoodoo option in the new win10fix settings.
Alternative: Update your AAU and see if it happens again. If so, disable win10fix, enable wined3d and software vertex processing.
>HELP! Required Windows 11 update broke things!
winkey+R -> ms-settings:developers -> Terminal=Windows Console Host

Previous Thread:
>>>/vg/428858839

Anonymous

Anonymous Fri 19 May 2023 07:51:15 No.7383460 Report

Quoted By:

>>7382356
can you at least roughly demonstrate how this is done..

Anonymous

Anonymous Fri 19 May 2023 15:05:41 No.7383861 Report

Quoted By:

>>7379794
Top tier outfit

Anonymous

Anonymous Fri 19 May 2023 23:07:10 No.7384588 Report

Quoted By:

>>7379152
catbox please? that's such a nice ass

Anonymous

Anonymous Sun 21 May 2023 23:47:11 No.7388582 Report

Quoted By:

>>7379351
>>7379593
>>7379675
>>7379721
>>7379743
>>7380705
>>7380743
>>7382058
>>7382113
>>7382364

Apologies for newfagging, relatively new to genning.
Does the character have a name or is somewhere in the LORAs listed in OP? Or is it an OC Lora or OC made by just prompting?

or am i too retarded to understand how something related to it can be extracted from the catbox pngs

Appreciate answers

Subject

Name

E-mail

Password

Capcode	All Only User Posts Only Verified Posts Only Moderator Posts Only Manager Posts Only Admin Posts Only Developer Posts Only Founder Posts
Show Posts	All Only With Images Only Without Images Only Spoiler Images Only Non-Spoiler Images
Deleted Posts	All Only Deleted Posts Only Non-Deleted Posts
Ghost Posts	All Only Ghost Posts Only Non-Ghost Posts
Post Type	All Only Sticky Threads Only Opening Posts Only Reply Posts
Results	All Grouped By Threads
Order	Latest Posts First Oldest Posts First

On these archives

On these boards

Your latest searches

/hdg/ - Hentai Diffusion General #323