/h/ - Hentai » Thread #7386103

>>7386142
>lora abuse
>needs a whole ass lora
Still retarded then. The entire point is modularity. I don't want one giant fuckoff cover all model since what I want is highly localised changes.
Idiotic reasoning that fundamentally misunderstands how AI image gen works.
These sorts of retards press one button and call it done. If you're not inpainting you're not producing the best shit.

Anonymous

View Same Google ImgOps iqdb SauceNAO Virtuous Contract.jpg, 655KiB, 2048x3072

Anonymous Sat 20 May 2023 21:18:44 No.7386151 Report

Quoted By: >>7386155 >>7386185

Can't remember if this one ever made it to /hdg/.

Anonymous

Anonymous Sat 20 May 2023 21:19:39 No.7386155 Report

Quoted By:

>>7386142
They might be furries, but at least they aren't (You)
>>7386151
I member

Anonymous

Anonymous Sat 20 May 2023 21:20:20 No.7386156 Report

Quoted By:

>>7386137
>>7386143
I gave a few comments in the last thread, but if I'm going to be picky about this specifically, the catbox userscript link should go to the original rentry I set up incase I ever need to migrate it off gist for some reason. Uses less chars than the gist link too. https://rentry.org/hdgcb

Anonymous

View Same Google ImgOps iqdb SauceNAO 2023-05-20_16-22-37.png, 860KiB, 852x519

Anonymous Sat 20 May 2023 21:24:03 No.7386164 Report

Quoted By: >>7386177

>>7386145
I guess I'll say it, the models they're hyping (fluffyrock and fluffusion) are shit
>looking at the thread
damn their shit is more fried than civitai

Anonymous

Anonymous Sat 20 May 2023 21:24:40 No.7386169 Report

Quoted By: >>7386172

>>7386146
>The setup links are a bit verbose and could be limited to 1 just for the wiki but it does help with discoverability
God, I hate you people.

Anonymous

View Same Google ImgOps iqdb SauceNAO curves text.jpg, 294KiB, 1024x1536

Anonymous Sat 20 May 2023 21:25:03 No.7386170 Report

Quoted By: >>7386185 >>7391163

so we're just using the borked thread? aight

Anonymous

Anonymous Sat 20 May 2023 21:26:25 No.7386172 Report

Quoted By: >>7386180

>>7386169
Yeah, and this is why I don't bake anymore.

Anonymous

Anonymous Sat 20 May 2023 21:26:37 No.7386173 Report

Quoted By: >>7386176

>>7386149
You're right that's why I use a lora for every pose, a lora for every hairstyle, a lora for every background

Anonymous

Anonymous Sat 20 May 2023 21:28:08 No.7386176 Report

Quoted By: >>7386189

>>7386173
You say that ironically but you're correct. I want the base model to be as neutral as possible.

Anonymous

Anonymous Sat 20 May 2023 21:28:30 No.7386177 Report

Quoted By: >>7386178

>>7386164
This is like digging into our previous thread and going "Look, /h/ is worse than civitai! >>7382711 "

Anonymous

Anonymous Sat 20 May 2023 21:29:05 No.7386178 Report

Quoted By:

>>7386177
>Look, /h/ is worse than civitai
I mean yeah

Anonymous

View Same Google ImgOps iqdb SauceNAO catbox_1fgzdh.png, 1MiB, 1024x1280

Anonymous Sat 20 May 2023 21:29:29 No.7386179 Report

Quoted By:

Anonymous

Anonymous Sat 20 May 2023 21:30:05 No.7386180 Report

Quoted By:

>>7386172
Shit's ridiculous.

Anonymous

Anonymous Sat 20 May 2023 21:30:30 No.7386181 Report

Quoted By: >>7386187

so which is the real one?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00234-1383465528-nsfw, scan, (tr (...).png, 2MiB, 1024x1536

Anonymous Sat 20 May 2023 21:30:37 No.7386182 Report

Quoted By: >>7386184 >>7386215 >>7386235 >>7386258 >>7386324

THREAD THEME

Anonymous

Anonymous Sat 20 May 2023 21:32:04 No.7386184 Report

Quoted By:

>>7386182
that's some good shit right there

Anonymous

Anonymous Sat 20 May 2023 21:32:10 No.7386185 Report

Quoted By: >>7386191

>>7386151
I think it didnt, coolio.

>>7386170
seems so. If you're the anon asking about english then yes they dont mind captions like this afaik. If you wanna please everyone you can just make an alt with it google translated to jp lol

Anonymous

Anonymous Sat 20 May 2023 21:33:05 No.7386187 Report

Quoted By:

>>7386181
I'm the real one, shoot him!

Anonymous

Anonymous Sat 20 May 2023 21:33:45 No.7386189 Report

Quoted By:

>>7386176
based SD1.5 user

Anonymous

Anonymous Sat 20 May 2023 21:34:21 No.7386191 Report

Quoted By: >>7386201

>>7386185
you mean, chatgpt translation. The only problem is giving it context without being flagged nor gpt caving in due to forbidden content. I wish there was a non judgemental translation focused one already

Anonymous

View Same Google ImgOps iqdb SauceNAO 00037-2962893417.png, 901KiB, 1024x1536

Anonymous Sat 20 May 2023 21:34:29 No.7386192 Report

Quoted By: >>7386208

Anonymous

View Same Google ImgOps iqdb SauceNAO 00455.jpg, 454KiB, 2048x3072

Anonymous Sat 20 May 2023 21:35:40 No.7386193 Report

Quoted By: >>7386202 >>7392268

On the topic of Pixiv vs AIBooru, there is one glaring flaw of Pixiv in that you cannot share uncensored images. I've got a bunch of those locally, any filehosts that won't trigger takedowns?

Anonymous

View Same Google ImgOps iqdb SauceNAO 9996490989765778-1230798768-kill (...).png, 3MiB, 1800x2400

Anonymous Sat 20 May 2023 21:36:16 No.7386195 Report

Quoted By: >>7386198

hopefully killjoy enjoyers like peaches because im going to post few before going to sleep

Anonymous

Anonymous Sat 20 May 2023 21:36:37 No.7386196 Report

Quoted By: >>7386199 >>7386204 >>7386206

>>7386149
i can get "modularity" part if you're using character lora because if your character isn't hatsune miku you'll most likely end up with some random girl instead of your prompted waifu
but a fucking doggystyle lora? a fucking paizuri lora? is it THAT bad?

Anonymous

View Same Google ImgOps iqdb SauceNAO 9996490989765796-3834760677-kill (...).png, 2MiB, 1600x1600

Anonymous Sat 20 May 2023 21:37:08 No.7386198 Report

Quoted By: >>7386200

>>7386195

Anonymous

Anonymous Sat 20 May 2023 21:37:26 No.7386199 Report

Quoted By: >>7386203

>>7386196
go back already
PLEASEEE

Anonymous

View Same Google ImgOps iqdb SauceNAO 9996490989765851-1302091150-kill (...).png, 2MiB, 2000x1800

Anonymous Sat 20 May 2023 21:37:41 No.7386200 Report

Quoted By:

>>7386198

Anonymous

Anonymous Sat 20 May 2023 21:37:52 No.7386201 Report

Quoted By: >>7386213

>>7386191
I wouldnt trust chatgpt but I havent tried

Anonymous

Anonymous Sat 20 May 2023 21:38:56 No.7386202 Report

Quoted By:

>>7386193
just do a light censor or toss a single bar across it.
tho I guess its painful if ur doing like 20 imgs

Anonymous

Anonymous Sat 20 May 2023 21:39:31 No.7386203 Report

Quoted By: >>7386211

>>7386199
but im already home and previous bread is exhausted, genning plot for shit i trained atm

Anonymous

Anonymous Sat 20 May 2023 21:39:58 No.7386204 Report

Quoted By: >>7386214

>>7386196
if your dogshit model was any good it would have been merged with others as an improvement, fuck off already back to your dogshit thread

Anonymous

View Same Google ImgOps iqdb SauceNAO 73244-1088956746-solo, 1girl,_lu (...).png, 811KiB, 640x1280

Anonymous Sat 20 May 2023 21:40:44 No.7386206 Report

Quoted By: >>7386217 >>7386231

>>7386196
you want doggystyle arched back? legs up holding? you want it standing or on all fours?
you want the angle POV or from above?
what about holding the tits from behind, or holding waist?
Yeah I really don't mind introducing multiple single purpose loras because what I want to see is often extremely specific. I don't want someone else's finetune version of what I have in my head. If you just want to roll some gacha and happily accept whatever shows up then yeah a more tuned, highly opinionated model is of course a better option.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00452-385510756.png, 843KiB, 1024x1536

Anonymous Sat 20 May 2023 21:40:47 No.7386208 Report

Quoted By:

>>7386192

Anonymous

Anonymous Sat 20 May 2023 21:41:34 No.7386211 Report

Quoted By:

>>7386203
Then do something about it yourself if it so bothers you instead of complaining for the past month, idiot.

Anonymous

Anonymous Sat 20 May 2023 21:42:06 No.7386213 Report

Quoted By:

>>7386201
it's an amazing translator, way better than google translate, but it stops the moment you mention nsfw

Anonymous

Anonymous Sat 20 May 2023 21:42:06 No.7386214 Report

Quoted By:

>>7386204
why you're calling this thread dogshit, literally friendlyfire

Anonymous

View Same Google ImgOps iqdb SauceNAO 00076-3826941901.png, 1MiB, 1024x1536

Anonymous Sat 20 May 2023 21:42:57 No.7386215 Report

Quoted By: >>7386223 >>7386229

>>7386182
Cat cutout works surprisingly well out of the box.

Anonymous

View Same Google ImgOps iqdb SauceNAO 61692-3256867576-diip, 1boy, kne (...).png, 904KiB, 768x1024

Anonymous Sat 20 May 2023 21:43:42 No.7386217 Report

Quoted By: >>7386231 >>7386253

>>7386206
Or view this from the other side. If your model doesn't inherently support what you want to see, aren't you forced to use loras anyway? I run into limits of the model constantly but it doesn't bother me because I extend the model's understanding effortlessly and on-demand. How can this ever be viewed as inferior to trying to cram everyone's ideals into one model, too many cooks.

Anonymous

Anonymous Sat 20 May 2023 21:45:24 No.7386219 Report

Quoted By: >>7386253

1girl lora when

Anonymous

Anonymous Sat 20 May 2023 21:47:03 No.7386223 Report

Quoted By: >>7386271

>>7386215
love your gens

what about the pussy peek, any tips? When I use the pussy peek tag it peeks from the sides but I want it super lowleg panties like this pic

Anonymous

View Same Google ImgOps iqdb SauceNAO 00211-2973659994-), blonde hair, (...).png, 2MiB, 1280x1536

Anonymous Sat 20 May 2023 21:47:09 No.7386224 Report

Quoted By:

anons, if you use wildcards/dynamic prompts, since an update from 3 days ago on dev (not master), they're broken, in the sense it uses info of any of the possible values inside the {} brackets instead of just 1.

https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/10565

Using older commit mentioned in the issue, being on master, or the vlad fork works as it should.

Anonymous

Anonymous Sat 20 May 2023 21:47:58 No.7386226 Report

Quoted By: >>7386231 >>7386232 >>7386253

So is ai art just better than normal art at this point, is there really a use for artists then?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00100-1228202600-nsfw, scan, (tr (...).png, 2MiB, 1024x1536

Anonymous Sat 20 May 2023 21:48:52 No.7386229 Report

Quoted By:

>>7386215
Yepp I'm surprised how well it works also nice gen.

Anonymous

Anonymous Sat 20 May 2023 21:49:10 No.7386231 Report

Quoted By: >>7386253

>>7386206
>>7386217
i guess im less weeb than average anon in this thread, good to know

>>7386226
you need a better bait

Anonymous

Anonymous Sat 20 May 2023 21:49:31 No.7386232 Report

Quoted By:

>>7386226
Oh yeah, making training data for new characters, obviously.

Anonymous

View Same Google ImgOps iqdb SauceNAO 01005-1104510590.png, 1MiB, 1024x1536

Anonymous Sat 20 May 2023 21:51:12 No.7386235 Report

Quoted By: >>7386242 >>7386258

>>7386182

Anonymous

View Same Google ImgOps iqdb SauceNAO beauty.png, 3MiB, 1433x2150

Anonymous Sat 20 May 2023 21:51:38 No.7386236 Report

Quoted By: >>7386243

I love my wife

Anonymous

Anonymous Sat 20 May 2023 21:55:09 No.7386242 Report

Quoted By: >>7386258

>>7386235
catbox now plz

Anonymous

Anonymous Sat 20 May 2023 21:56:50 No.7386243 Report

Quoted By: >>7386278

>>7386236
We've all "loved" your wife once or twice too.

Anonymous

View Same Google ImgOps iqdb SauceNAO 13345-2224835401-_closed eyes, w (...).png, 2MiB, 1152x1536

Anonymous Sat 20 May 2023 22:00:48 No.7386253 Report

Quoted By: >>7386266

>>7386217
I remember training the abmono textual embedding, hypernetwork, dreambooth and lora probably in that order.
I remember using the dreambooth finetune with different TIs, hypernets and with loras and everything got pretty fucked up because it was overfit on one style. It did that style fantastically well but because it was so opinionated, it ruined the flexibility of the model.

>>7386219
again I get that you're ironic but there's merit to that. being able to seamlessly blend between different variants of characters is pretty much that. you finetune the model's understanding of "1girl"
each character, style, etc. In the same way that cockdealer anon sperged out for months over his useless mixes, there was never any practical value to trying to share 2GB+ files. I know this, I trained and shared dreambooths and very few people gave a shit, most people did not care.
once I dropped the same thing as a bunch of loras when kohya became popularised, the tone drastically changed. TI anon was still assblasted but there are always contrarians.

>>7386231
you're in a hentai thread?
do you go to a furry thread and guess if your'e less furry than average?

>>7386226
training data, mutliple character interactions, complex poses, new designs, on-model designs, hands, coherence, sensible backgrounds
mostly training data
without regular artists there would be no submissive kronii with giant tits and a facial

Anonymous

View Same Google ImgOps iqdb SauceNAO 01009-692246630.png, 450KiB, 512x768

Anonymous Sat 20 May 2023 22:02:05 No.7386258 Report

Quoted By: >>7386261

>>7386182
>>7386235

>>7386242
https://litter.catbox.moe/lmgopw.png
https://litter.catbox.moe/nmmf7a.png

Anonymous

View Same Google ImgOps iqdb SauceNAO 1662421351924564.png, 2MiB, 1280x1600

Anonymous Sat 20 May 2023 22:03:32 No.7386260 Report

Quoted By: >>7386280 >>7389132

One quick gen before sleep. The cat cutout works pretty neat, but I had to put animal ears in the negatives for obvious reasons.

Anonymous

Anonymous Sat 20 May 2023 22:04:20 No.7386261 Report

Quoted By:

>>7386258
thank you brotheren

Anonymous

Anonymous Sat 20 May 2023 22:06:10 No.7386266 Report

Quoted By:

>>7386253
its scarier to visit 4chan furry threads than sift through pixiv posts while gathering dataset for obscure character

Anonymous

Anonymous Sat 20 May 2023 22:07:28 No.7386269 Report

Quoted By: >>7386286

>>7386137
i can actually appreciate this op. too bad the faggot that keeps spamming his shitted gutted one is just going to keep using his

Anonymous

Anonymous Sat 20 May 2023 22:07:54 No.7386271 Report

Quoted By:

>>7386223
Thanks but that was gacha. Sorry I don't got any tips for pussy peak. The AI will do it for itself sometimes though.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00242-3572944801-best quality,ns (...).png, 2MiB, 1280x1536

Anonymous Sat 20 May 2023 22:14:16 No.7386276 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO afterschoolselfie.png, 4MiB, 1383x1944

Anonymous Sat 20 May 2023 22:14:49 No.7386278 Report

Quoted By:

>>7386243
Hot

Anonymous

Anonymous Sat 20 May 2023 22:15:51 No.7386280 Report

Quoted By:

>>7386260
how do anons get their gens to look so sharp?

Anonymous

View Same Google ImgOps iqdb SauceNAO 01136.png, 2MiB, 1539x1524

Anonymous Sat 20 May 2023 22:16:54 No.7386284 Report

Quoted By: >>7386287 >>7386294

Is there any benefit to include backgrounds from a CG set into a style? Probably would spoil the whole thing right?

Anonymous

Anonymous Sat 20 May 2023 22:17:33 No.7386286 Report

Quoted By:

>>7386269
I don't think he posted two threads at once himself.

Anonymous

Anonymous Sat 20 May 2023 22:18:12 No.7386287 Report

Quoted By: >>7386390

>>7386284
like just backgrounds without any characters? probably not, unless you want the backgrounds to look like that

Anonymous

View Same Google ImgOps iqdb SauceNAO spurt.jpg, 265KiB, 1024x1536

Anonymous Sat 20 May 2023 22:20:36 No.7386290 Report

Quoted By: >>7386301

Anonymous

Anonymous Sat 20 May 2023 22:21:37 No.7386294 Report

Quoted By: >>7386390

>>7386284
You baking a Shamsiel lora?

Anonymous

Anonymous Sat 20 May 2023 22:24:32 No.7386301 Report

Quoted By: >>7386303

>>7386290
What?

Anonymous

Anonymous Sat 20 May 2023 22:25:43 No.7386303 Report

Quoted By: >>7386305

>>7386301
what what

Anonymous

Anonymous Sat 20 May 2023 22:27:42 No.7386305 Report

Quoted By: >>7386307 >>7386315 >>7386318

>>7386303
I don't get the dialogue. Why would he tell his mom?

Anonymous

Anonymous Sat 20 May 2023 22:28:05 No.7386307 Report

Quoted By: >>7386318

>>7386305
to make the mother jealous

Anonymous

Anonymous Sat 20 May 2023 22:31:04 No.7386315 Report

Quoted By: >>7386320

>>7386305
she's having a clandestine relationship with a shota and he innocently may think that painting over his babysitter/big sis/cousin is an ok thing

Anonymous

View Same Google ImgOps iqdb SauceNAO xyz_grid-0020-1756921160-solo, N (...).jpg, 346KiB, 4000x409

Anonymous Sat 20 May 2023 22:31:10 No.7386316 Report

Quoted By: >>7386579

>delete the image with stairs
>rebake
for what reason does this continue, what the actual fuck. The only difference is tagger strength, I just dont fucking understand.
>doesnt cause problems when the prompt is "walking"
whatever man fuck this I give up, there's nothing with the "sitting" tag that could be causing this in particular, especially tag-wise.

Anonymous

Anonymous Sat 20 May 2023 22:32:14 No.7386318 Report

Quoted By:

>>7386305
>>7386307
or this, both are milking him and each think they're exclusive, use your imagination bruv, fill in the gaps

Anonymous

Anonymous Sat 20 May 2023 22:33:57 No.7386320 Report

Quoted By: >>7392271

>>7386315
Oh, I didn't notice the shota dick

Anonymous

Anonymous Sat 20 May 2023 22:34:18 No.7386321 Report

Quoted By:

>>7385867
Catboxu again?

Anonymous

View Same Google ImgOps iqdb SauceNAO 44778-2044639848.0-solo, pink ha (...).png, 3MiB, 1280x1792

Anonymous Sat 20 May 2023 22:37:04 No.7386324 Report

Quoted By: >>7386326 >>7386609

>>7386182
Sure why not

Anonymous

Anonymous Sat 20 May 2023 22:38:16 No.7386326 Report

Quoted By:

>>7386324
>look at me mom, all 10 fingers!

Anonymous

View Same Google ImgOps iqdb SauceNAO catbox_bgaofl.png, 1MiB, 1024x1280

Anonymous Sat 20 May 2023 22:39:37 No.7386332 Report

Quoted By: >>7386399

Anonymous

Anonymous Sat 20 May 2023 22:44:29 No.7386344 Report

Quoted By: >>7386346

Anyone have experience making anime screenshot LORAs here? I want to know the best method for collecting frames

Anonymous

Anonymous Sat 20 May 2023 22:45:21 No.7386346 Report

Quoted By: >>7386373

>>7386344
Manual or run FFMPEG at low framerates. Czkawka later.

Anonymous

Anonymous Sat 20 May 2023 22:57:50 No.7386367 Report

Quoted By: >>7386378

so i cant continue training after resizing lora, get dim mismatch even if i set it to the same, am i missing a step?

Anonymous

Anonymous Sat 20 May 2023 23:01:37 No.7386373 Report

Quoted By: >>7386386

>>7386346
>Czkawka later.
is there a way to use FFMPEG within python and receiving the frames as numpy arrays so i could use my own hashing algos?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00087-3017444038.png, 2MiB, 1024x1536

Anonymous Sat 20 May 2023 23:02:05 No.7386374 Report

Quoted By:

Anonymous

Anonymous Sat 20 May 2023 23:03:15 No.7386378 Report

Quoted By: >>7386380

>>7386367
yes

Anonymous

Anonymous Sat 20 May 2023 23:05:50 No.7386380 Report

Quoted By: >>7386387

>>7386378
and that is?

Anonymous

View Same Google ImgOps iqdb SauceNAO Wallbang.jpg, 245KiB, 1024x1792

Anonymous Sat 20 May 2023 23:08:16 No.7386385 Report

Quoted By: >>7386395 >>7386579

Anonymous

Anonymous Sat 20 May 2023 23:08:27 No.7386386 Report

Quoted By:

>>7386373
I dunno. Seems to me like you're the one that should know, heh.

Anonymous

Anonymous Sat 20 May 2023 23:09:04 No.7386387 Report

Quoted By:

>>7386380
--dim_from_weights

Anonymous

View Same Google ImgOps iqdb SauceNAO 01137.png, 2MiB, 1353x1483

Anonymous Sat 20 May 2023 23:11:12 No.7386390 Report

Quoted By:

>>7386287
I thought so, they look alright at the moment with the right mix, even with just a character lora.
>>7386294
Already did, thinking about making a style lora from the artist.

Anonymous

Anonymous Sat 20 May 2023 23:13:31 No.7386395 Report

Quoted By:

>>7386385
you must be at least this tall to ride

Anonymous

Anonymous Sat 20 May 2023 23:14:20 No.7386399 Report

Quoted By: >>7386428

>>7386332
did you make a bebebe lora?

Anonymous

Anonymous Sat 20 May 2023 23:14:43 No.7386400 Report

Quoted By:

Is raising max token length from 150 to 300 fine or am I gonna be running into some issues caused by this? I mean for training, not prompts.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00311-1391140534.png, 2MiB, 1280x1536

Anonymous Sat 20 May 2023 23:18:11 No.7386404 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO catbox_qdeuf5.png, 1MiB, 1024x1280

Anonymous Sat 20 May 2023 23:28:22 No.7386428 Report

Quoted By: >>7387820

>>7386399
Yes, just finished baking it, and enjoying the results so far

Anonymous

Anonymous Sat 20 May 2023 23:31:41 No.7386435 Report

Quoted By: >>7386450 >>7386460 >>7386479

question at the risk of sounding immensely retarded
people always say not to pad out your dataset with AI genned images, but why is this? i know badly genned images can fuck up your dataset but what if there are no noticable flaws in the ones used

Anonymous

View Same Google ImgOps iqdb SauceNAO 02744-2897378296-masterpiece, be (...).png, 2MiB, 1536x2400

Anonymous Sat 20 May 2023 23:36:35 No.7386446 Report

Quoted By: >>7386456 >>7386612

she forgot something

Anonymous

Anonymous Sat 20 May 2023 23:37:14 No.7386450 Report

Quoted By: >>7386458

>>7386435
Granted, I only did this once and months ago, but negative feedback loop.
You're training the AI to reinforce its weaknesses.

Anonymous

Anonymous Sat 20 May 2023 23:37:58 No.7386453 Report

Quoted By:

Does anyone saved seeds from old decent gens with flawed prompts just in case they might revisit them?

Anonymous

Anonymous Sat 20 May 2023 23:39:12 No.7386456 Report

Quoted By:

>>7386446
yeah, razor

Anonymous

Anonymous Sat 20 May 2023 23:39:26 No.7386458 Report

Quoted By:

>>7386450
makes sense
sucks because i'm having a hard time finding suitable art for some characters
had to remove some official art for a character lora too because it forced a super rendered look onto generations

Anonymous

Anonymous Sat 20 May 2023 23:40:34 No.7386460 Report

Quoted By:

>>7386435
It's only worth it if your dataset is REALLY small like <10 images. You'll have to make sure they don't have fucked up backgrounds or other artifacts, and use multiple models/style to avoid overfitting style.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00352-3354803297.png, 2MiB, 1280x1536

Anonymous Sat 20 May 2023 23:46:52 No.7386469 Report

Quoted By: >>7389544

Anonymous

View Same Google ImgOps iqdb SauceNAO Natsume5_001.jpg, 85KiB, 960x1361

Anonymous Sat 20 May 2023 23:50:32 No.7386477 Report

Quoted By:

man I really wanna keep this pic just for the condoms, but I think it might be causing random feet showing up. I doubt just tagging it "leg up" would circumvent that?

Anonymous

Anonymous Sat 20 May 2023 23:52:06 No.7386479 Report

Quoted By:

>>7386435
because there will be stray hair or fucked up pixels that you missed, AI will learn them and keep re-doing these fuckups. I trained a decent lora on 10 images so I dont think you'll ever have to rely on that

Anonymous

View Same Google ImgOps iqdb SauceNAO 1657821511529.png, 1MiB, 1280x1024

Anonymous Sat 20 May 2023 23:54:04 No.7386484 Report

Quoted By: >>7386490 >>7387726

controlnet reference seems magical

Anonymous

Anonymous Sat 20 May 2023 23:54:26 No.7386486 Report

Quoted By: >>7386493

So is WD1.5 any good?

Anonymous

View Same Google ImgOps iqdb SauceNAO 1672063589865.png, 2MiB, 1280x1024

Anonymous Sat 20 May 2023 23:54:41 No.7386487 Report

Quoted By:

Anonymous

Anonymous Sat 20 May 2023 23:56:15 No.7386490 Report

Quoted By: >>7386504

>>7386484
what was the OG? im still sticking to canny and hed

Anonymous

Anonymous Sat 20 May 2023 23:57:03 No.7386493 Report

Quoted By:

>>7386486
no it's beyond fucked
all the variant models are also fried to shit and can barely gen any variety without high weight prompting

Anonymous

View Same Google ImgOps iqdb SauceNAO 1673377769911.png, 1MiB, 1920x512

Anonymous Sun 21 May 2023 00:03:19 No.7386504 Report

Quoted By: >>7386514 >>7386526 >>7387267

>>7386490
the left + middle = right

Anonymous

View Same Google ImgOps iqdb SauceNAO guest jpg.jpg, 258KiB, 1024x1536

Anonymous Sun 21 May 2023 00:05:23 No.7386508 Report

Quoted By:

I need thick smooth thighs to jizz between to

Anonymous

Anonymous Sun 21 May 2023 00:08:18 No.7386514 Report

Quoted By: >>7386528

>>7386504
I dont get it

Anonymous

View Same Google ImgOps iqdb SauceNAO 00373-3446374313.png, 2MiB, 1280x1536

Anonymous Sun 21 May 2023 00:08:46 No.7386516 Report

Quoted By: >>7386612

Anonymous

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

Anonymous Sun 21 May 2023 00:09:21 No.7386519 Report

Quoted By: >>7386534

Anonymous

Anonymous Sun 21 May 2023 00:10:08 No.7386523 Report

Quoted By: >>7386611

>>7385806
Man, I just don't get how you get lines that crisp with those settings.
Is it the mix?

Anonymous

Anonymous Sun 21 May 2023 00:11:14 No.7386525 Report

Quoted By:

>>7385933
they're either dumb or purposefully misunderstanding you. i second the request for a lora that specializes in a variety of highly detailed underwear

>>7386040
use either tiled diffusion or controlnet tile sampler + tile model along side ultimate SD

>>7386055]
having a view that groups search results by author under one preview image would help. the preview image could be hovered to reveal more images from said author matching said tag. throw in a rating on that authors preview image so you can skip spammers
2540# "Post menu")

>>7386068
moralfags have no business modding a booru. unless it's illegal or clearly to harrass a real person with fake AI art, you can kindly stfu with your thoughts on xyz kink

Anonymous

Anonymous Sun 21 May 2023 00:11:19 No.7386526 Report

Quoted By: >>7386528

>>7386504
Didn't know you can combine images

Anonymous

Anonymous Sun 21 May 2023 00:12:12 No.7386528 Report

Quoted By:

>>7386526
>>7386514
https://github.com/Mikubill/sd-webui-controlnet/discussions/1236

Anonymous

Anonymous Sun 21 May 2023 00:16:54 No.7386533 Report

Quoted By:

>>7386068
>filtering non-con
what kind of weak shit is this

Anonymous

View Same Google ImgOps iqdb SauceNAO catbox_m2jzn3.png, 1MiB, 1024x1024

Anonymous Sun 21 May 2023 00:16:55 No.7386534 Report

Quoted By: >>7386541

>>7386519

Anonymous

Anonymous Sun 21 May 2023 00:22:12 No.7386541 Report

Quoted By: >>7386545

>>7386534
What's that model?

Anonymous

View Same Google ImgOps iqdb SauceNAO mommy.png, 1MiB, 896x1600

Anonymous Sun 21 May 2023 00:22:46 No.7386542 Report

Quoted By: >>7386612

Anonymous

View Same Google ImgOps iqdb SauceNAO catbox_2hid80.png, 1MiB, 1024x1024

Anonymous Sun 21 May 2023 00:24:06 No.7386545 Report

Quoted By:

>>7386541
Mokumix

Anonymous

View Same Google ImgOps iqdb SauceNAO 00114-3759464760.png, 2MiB, 1024x1536

Anonymous Sun 21 May 2023 00:26:41 No.7386547 Report

Quoted By: >>7386549

Anonymous

Anonymous Sun 21 May 2023 00:27:46 No.7386549 Report

Quoted By:

>>7386547
me on the left

Anonymous

Anonymous Sun 21 May 2023 00:27:48 No.7386550 Report

Quoted By:

>>7386103
Last time I frequented these threads, there was a big controlnet update that seemed promising. Has there been any other major updates since then? Also has anyone found an alternate to colab now that theyve started cracking down on us?

Anonymous

View Same Google ImgOps iqdb SauceNAO catbox_enc1bc.png, 1MiB, 1024x1024

Anonymous Sun 21 May 2023 00:36:00 No.7386562 Report

Quoted By:

Anonymous

Anonymous Sun 21 May 2023 00:37:01 No.7386563 Report

Quoted By: >>7386575

newish noob here: i'm using a lora with a preexisting image while using controlnet softedge, but all of the outputs are weird and super bright/washed out (it doesn't look like a vae issue where it's gray/dull, but if someone knows how to make sure vae was used on an image, lmk and i'll double check).

are concept loras just not meant to be used with controlnet? i'm thinking about testing using the controlnet first, then running that through img2img with the lora activated, but images take a while to finish on my low end pc

Anonymous

Anonymous Sun 21 May 2023 00:40:18 No.7386571 Report

Quoted By:

Ok, bad-hands is pretty terrible, but that add detail lora is interesting...

Anonymous

Anonymous Sun 21 May 2023 00:42:03 No.7386575 Report

Quoted By:

>>7386563
did you try a different controlnet processor/preprocessor? shit is complicated. if it's a concept you can apply later, i'd just inpaint it in

Anonymous

Anonymous Sun 21 May 2023 00:43:42 No.7386579 Report

Quoted By: >>7386628

>>7385076
i sloppily drew some green stuff over the plate and inpainted it locally with "slime, green slime on plate" in prompt, that was first try too, i was surprised it got what i meant so quickly. was too tired to make it smoother

>>7386316
that's so weird. it might have nothing to do with the training but i wonder if it would persist on other mixes

>>7386385
tastefully half-exposed tits, based. is that a font with manually added paint streaks?

Anonymous

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

Anonymous Sun 21 May 2023 00:45:00 No.7386583 Report

Quoted By: >>7392288

what the fuck

Anonymous

Anonymous Sun 21 May 2023 00:47:59 No.7386589 Report

Quoted By: >>7386590 >>7387134

Glad I didn't invest in a 4090, things really did plateau since January.

Anonymous

Anonymous Sun 21 May 2023 00:48:50 No.7386590 Report

Quoted By: >>7387134

>>7386589
>4000 series
lmao even

Anonymous

Anonymous Sun 21 May 2023 00:53:12 No.7386596 Report

Quoted By: >>7386597

That reminds me do ray tracing/tensor cores matter or is it just only CUDA cores

Anonymous

Anonymous Sun 21 May 2023 00:53:58 No.7386597 Report

Quoted By:

>>7386596
I'd argue memory is more important here but for speed yeah, CUDA.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00259-858025093-nsfw, scan, (tra (...).png, 2MiB, 1024x1536

Anonymous Sun 21 May 2023 00:58:34 No.7386599 Report

Quoted By: >>7386600 >>7386612

Anonymous

Anonymous Sun 21 May 2023 01:00:18 No.7386600 Report

Quoted By:

>>7386599
gato

Anonymous

View Same Google ImgOps iqdb SauceNAO 00413-2288811305.png, 2MiB, 1280x1408

Anonymous Sun 21 May 2023 01:09:35 No.7386606 Report

Quoted By: >>7386610

Anonymous

Anonymous Sun 21 May 2023 01:12:21 No.7386609 Report

Quoted By: >>7386729

>>7386324
Your hand looks so good how you make it?

Anonymous

Anonymous Sun 21 May 2023 01:13:37 No.7386610 Report

Quoted By: >>7386614

>>7386606
gato

Anonymous

View Same Google ImgOps iqdb SauceNAO 01112-2564502437, DPM++ 2M Karra (...).jpg, 493KiB, 768x1280

Anonymous Sun 21 May 2023 01:13:40 No.7386611 Report

Quoted By: >>7386615 >>7386631

>>7386523
probably

Anonymous

Anonymous Sun 21 May 2023 01:13:41 No.7386612 Report

Quoted By: >>7386614 >>7386624 >>7387280

>>7386446
>>7386516
>>7386542
>>7386599
Catbox please?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00416-3616007975.png, 2MiB, 1280x1408

Anonymous Sun 21 May 2023 01:16:42 No.7386614 Report

Quoted By: >>7386615 >>7386620 >>7386622

>>7386610
https://files.catbox.moe/7xkkz9.png

>>7386612
https://files.catbox.moe/92m3w9.png

Anonymous

Anonymous Sun 21 May 2023 01:17:21 No.7386615 Report

Quoted By:

>>7386611
Yeah, testing it now and it's a big part.
Hate how the soulful mixes with thick styles fuck up the details tho.
>>7386614
TY

Anonymous

Anonymous Sun 21 May 2023 01:21:14 No.7386620 Report

Quoted By:

>>7386614
cute layler

Anonymous

Anonymous Sun 21 May 2023 01:23:38 No.7386622 Report

Quoted By: >>7386667

>>7386614
Anon, wth is that mix? Link?

Anonymous

Anonymous Sun 21 May 2023 01:24:31 No.7386624 Report

Quoted By: >>7386626 >>7386636

>>7386612
I'm away from PC now but for knight lady I used my messy pubes lora(its on gayshit) and the mommy is using the pepe style lora from civitai. model was based64

Anonymous

Anonymous Sun 21 May 2023 01:26:26 No.7386626 Report

Quoted By:

>>7386624
>messy pubes lora
BASED

Anonymous

Anonymous Sun 21 May 2023 01:27:31 No.7386628 Report

Quoted By:

>>7386579
>it might have nothing to do with the training but i wonder if it would persist on other mixes
I tested on aom2nutmeg and based64, its the lora not the models.
>i was surprised it got what i meant so quickly
didnt expect it to gen it that well either

Anonymous

Anonymous Sun 21 May 2023 01:28:11 No.7386631 Report

Quoted By: >>7386643

>>7386611
Did you not like v2 of yabal?

Anonymous

Anonymous Sun 21 May 2023 01:29:07 No.7386636 Report

Quoted By:

>>7386624
>messy pubes
hello KING

Anonymous

View Same Google ImgOps iqdb SauceNAO 00108-811041415, DPM++ 2M Karras (...).jpg, 535KiB, 768x1280

Anonymous Sun 21 May 2023 01:35:19 No.7386643 Report

Quoted By: >>7386645 >>7386658

>>7386631
didn't know there was a v2, found v1 through huggingface

Anonymous

Anonymous Sun 21 May 2023 01:35:49 No.7386645 Report

Quoted By:

>>7386643
Ah.
I'm looking at it right now. Want something stronger now.

Anonymous

Anonymous Sun 21 May 2023 01:37:54 No.7386650 Report

Quoted By: >>7386652

bruh scraping safebooru is so much slower than gelbooru

Anonymous

Anonymous Sun 21 May 2023 01:39:05 No.7386652 Report

Quoted By: >>7386654 >>7386658

>>7386650
Why are you even scraping safebooru?

Anonymous

Anonymous Sun 21 May 2023 01:41:19 No.7386654 Report

Quoted By: >>7386659

>>7386652
whats wrong with safebooru?

Anonymous

Anonymous Sun 21 May 2023 01:44:33 No.7386658 Report

Quoted By: >>7386659 >>7386671

>>7386652
I'm grabbing the top 2500 images with certain tags for an experiment; since danbooru's high rated images are all NSFW I need to get safebooru stuff too in order to ensure a non-NSFW-biased dataset.
>>7386643
There is a v3, even.

Anonymous

Anonymous Sun 21 May 2023 01:45:26 No.7386659 Report

Quoted By: >>7386662

>>7386654
Nothing, just thought it was deNSFWed danbooru.
>>7386658
Can't you just use a safe condition in grabber? 99.9% sure you can.

Anonymous

Anonymous Sun 21 May 2023 01:47:17 No.7386661 Report

Quoted By: >>7386663

Am I correct in assuming that each checkpoint or lora has an optimal aspect ratio that is based on its training, and if so, is it possible to override it? I don't know if this is what's happening, but the quality of results seem to vary depending on the combination of these factors, and it's driving me nuts.

Anonymous

Anonymous Sun 21 May 2023 01:47:29 No.7386662 Report

Quoted By:

>>7386659
Not using grabber. I could have used the rating: tags though. Oh well, the script's already running.

Anonymous

Anonymous Sun 21 May 2023 01:48:38 No.7386663 Report

Quoted By:

>>7386661
>is it possible to override it
No, not really? Unless you finetune it yourself, gacha, or force it through controlnet.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00446-2365800998.png, 2MiB, 1280x1408

Anonymous Sun 21 May 2023 01:52:54 No.7386667 Report

Quoted By: >>7386669 >>7386672

>>7386622
Self made mix (mixeranon), basically some models mixed with MBW with different weights, and then 50% of anime25like with alpha on 0. Haven't gave it a name.

You will get a similar result just doing 50% AOM2 and 50% anime25like I think.

I haven't uploaded it yet, just another with similar method but 50% of xxx model https://huggingface.co/BGLuck/BlockMergeWeightedMixes/blob/main/peculiar_xxmix.safetensors.

Anonymous

Anonymous Sun 21 May 2023 01:54:33 No.7386669 Report

Quoted By:

>>7386667
Oh. Too bad you didn't release it yet.

Anonymous

Anonymous Sun 21 May 2023 01:55:08 No.7386671 Report

Quoted By: >>7386678 >>7386679

>>7386658
where? couldn't find it in HF. Searching with five letters doesn't let me search in archive unfortunately.

Anonymous

Anonymous Sun 21 May 2023 01:55:18 No.7386672 Report

Quoted By: >>7386674

>>7386667
>animelike
nooo

Anonymous

View Same Google ImgOps iqdb SauceNAO 00448-3952717895.png, 2MiB, 1280x1408

Anonymous Sun 21 May 2023 01:55:59 No.7386674 Report

Quoted By:

>>7386672
:(

Anonymous

Anonymous Sun 21 May 2023 01:57:14 No.7386678 Report

Quoted By: >>7386681

>>7386671
on civit

Anonymous

Anonymous Sun 21 May 2023 01:57:25 No.7386679 Report

Quoted By: >>7386680 >>7386681

>>7386671
Civit. V3 seems to be the worse for me so far.

Anonymous

Anonymous Sun 21 May 2023 01:59:58 No.7386680 Report

Quoted By: >>7386710

>>7386679
Actually I don't know. Seems better in some places and worse in others.

Anonymous

Anonymous Sun 21 May 2023 02:00:52 No.7386681 Report

Quoted By:

>>7386678
>>7386679
Thanks. I probably mispelled when I searching for it in there.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00030-1192501403-1girl, mature f (...).png, 2MiB, 1280x1792

Anonymous Sun 21 May 2023 02:24:18 No.7386700 Report

Quoted By: >>7386710 >>7386714

Yabal's pretty nice.

Anonymous

View Same Google ImgOps iqdb SauceNAO 1684324769307001-569610967-from (...).png, 1MiB, 768x1152

Anonymous Sun 21 May 2023 02:31:50 No.7386706 Report

Quoted By: >>7386713

Anonymous

View Same Google ImgOps iqdb SauceNAO 1674639853386119.png, 847KiB, 1024x768

Anonymous Sun 21 May 2023 02:32:17 No.7386708 Report

Quoted By: >>7386713

Anonymous

View Same Google ImgOps iqdb SauceNAO 00242-719922-1girl, mature femal (...).png, 525KiB, 640x896

Anonymous Sun 21 May 2023 02:33:17 No.7386710 Report

Quoted By: >>7386716

>>7386680
>>7386700
Actually seems that v3 is the most anatomically consistent. I really like those.
Have to bake a bunch more grids tho. Later.
https://files.catbox.moe/c8o4ts.png
https://files.catbox.moe/pf4btc.png
https://files.catbox.moe/2kpab6.png

Anonymous

Anonymous Sun 21 May 2023 02:34:19 No.7386713 Report

Quoted By: >>7386719

>>7386706
>>7386708
gato?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00474-2233015579.png, 3MiB, 2560x1408

Anonymous Sun 21 May 2023 02:34:31 No.7386714 Report

Quoted By: >>7386716

>>7386700
I did a lana's mother the other day, didn't know there was other one out there :(

Anonymous

Anonymous Sun 21 May 2023 02:36:04 No.7386716 Report

Quoted By:

>>7386710
(It also did pov hands consistently well compared to the other ones.)
>>7386714
The more the merrier.

Anonymous

Anonymous Sun 21 May 2023 02:41:18 No.7386719 Report

Quoted By:

>>7386713
https://files.catbox.moe/0s1wj5.png
think i'm going to call it done for x-2 rikku it's reasonably accurate

Anonymous

Anonymous Sun 21 May 2023 02:42:21 No.7386722 Report

Quoted By: >>7386733

What are everyone's favourite style loras? Looking for something cute

Anonymous

View Same Google ImgOps iqdb SauceNAO 00486-2889204391.png, 2MiB, 1280x1408

Anonymous Sun 21 May 2023 02:49:19 No.7386728 Report

Quoted By: >>7386735

Anonymous

Anonymous Sun 21 May 2023 02:50:03 No.7386729 Report

Quoted By: >>7386800

>>7386609
Nothing, "waving" almost always come out good
catbox for reference
>https://files.catbox.moe/smq21z.png

Anonymous

Anonymous Sun 21 May 2023 02:52:39 No.7386733 Report

Quoted By:

>>7386722
Afrobull and Pokemon are pretty nice.

Anonymous

Anonymous Sun 21 May 2023 02:53:40 No.7386735 Report

Quoted By: >>7386751

>>7386728
gato

Anonymous

View Same Google ImgOps iqdb SauceNAO 1684324769307023-715602178-cowbo (...).png, 1MiB, 768x1152

Anonymous Sun 21 May 2023 02:56:29 No.7386737 Report

Quoted By:

Anonymous

Anonymous Sun 21 May 2023 02:58:40 No.7386739 Report

Quoted By:

>>7386103
a 6 fingered mutant as your op pic? really op?

Anonymous

View Same Google ImgOps iqdb SauceNAO 0-original.png, 1MiB, 768x1152

Anonymous Sun 21 May 2023 03:01:06 No.7386741 Report

Quoted By: >>7386744

Has anyone messed around with the Unstable Diffusion beta? It's better than base NAI but still a bit weak compared to mixes, though the local version isn't out yet. The samplers have some cool features like saturation control. Hopefully it can continue to improve so we finally escape the flaws of NAI mixes.

Anonymous

Anonymous Sun 21 May 2023 03:02:47 No.7386744 Report

Quoted By: >>7386747

>>7386741
gens on it look like shit and i'd rather just wait for an actual release before i start training on it

Anonymous

Anonymous Sun 21 May 2023 03:05:10 No.7386747 Report

Quoted By: >>7386754

>>7386744
I don't even think you can train on it yet? I think it's still a web-only thing for now.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00014-3757378905.png, 1MiB, 1024x1456

Anonymous Sun 21 May 2023 03:05:19 No.7386748 Report

Quoted By:

Girls in hoodies

Anonymous

View Same Google ImgOps iqdb SauceNAO 6102742813895.png, 2MiB, 864x1312

Anonymous Sun 21 May 2023 03:05:29 No.7386750 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO 00496-144428897.png, 1MiB, 1280x1408

Anonymous Sun 21 May 2023 03:06:01 No.7386751 Report

Quoted By:

>>7386735
https://files.catbox.moe/wqxief.png

Anonymous

Anonymous Sun 21 May 2023 03:07:37 No.7386754 Report

Quoted By: >>7387440

>>7386747
they released it for download a couple weeks ago i think? and suggested people train on it. someone brought it up in the thread

Anonymous

View Same Google ImgOps iqdb SauceNAO why can't anything be easy.png, 2MiB, 1024x1536

Anonymous Sun 21 May 2023 03:09:30 No.7386757 Report

Quoted By: >>7386759 >>7386761

How do I use the 840000 vae? I downloaded the vae-ft-mse-840000-ema-pruned.safetensors file and placed it in the folder where my other vaes are and it's not showing up in the dropdown.

Anonymous

View Same Google ImgOps iqdb SauceNAO 1664272097135284.png, 865KiB, 1024x768

Anonymous Sun 21 May 2023 03:09:39 No.7386758 Report

Quoted By:

Anonymous

Anonymous Sun 21 May 2023 03:11:54 No.7386759 Report

Quoted By: >>7386765

>>7386757
refresh

Anonymous

Anonymous Sun 21 May 2023 03:12:37 No.7386761 Report

Quoted By: >>7386765

>>7386757
sometimes u gotta restart webui for things to show up even though it technically should work on a restart

Anonymous

View Same Google ImgOps iqdb SauceNAO 1661992939732042.png, 438KiB, 512x768

Anonymous Sun 21 May 2023 03:20:32 No.7386764 Report

Quoted By:

Anonymous

Anonymous Sun 21 May 2023 03:21:22 No.7386765 Report

Quoted By:

>>7386759
>>7386761
Oh, awesome, thank you.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00020-426397622ed.png, 1MiB, 1024x1536

Anonymous Sun 21 May 2023 03:36:56 No.7386787 Report

Quoted By: >>7387159

Anonymous

View Same Google ImgOps iqdb SauceNAO 00024-2737149300.png, 1MiB, 1024x1456

Anonymous Sun 21 May 2023 03:49:11 No.7386799 Report

Quoted By:

Anonymous

Anonymous Sun 21 May 2023 03:49:13 No.7386800 Report

Quoted By:

>>7386729
>based waving tag enjoyer
not many of us

Anonymous

View Same Google ImgOps iqdb SauceNAO 00506-3250101119.png, 2MiB, 1280x1408

Anonymous Sun 21 May 2023 03:55:24 No.7386807 Report

Quoted By: >>7386840

Anonymous

View Same Google ImgOps iqdb SauceNAO 00532-3141599177.png, 1MiB, 1024x1536

Anonymous Sun 21 May 2023 04:09:03 No.7386821 Report

Quoted By: >>7386882 >>7387252

kirche has been oddly hard to train.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00007-640863325.png, 1MiB, 1024x1456

Anonymous Sun 21 May 2023 04:11:25 No.7386823 Report

Quoted By:

Anonymous

Anonymous Sun 21 May 2023 04:17:56 No.7386840 Report

Quoted By: >>7386845

>>7386807
catbox? I like the flat-ish colors

Anonymous

View Same Google ImgOps iqdb SauceNAO 00508-4173120661.png, 2MiB, 1280x1408

Anonymous Sun 21 May 2023 04:20:28 No.7386845 Report

Quoted By: >>7386851 >>7387170

>>7386840
https://files.catbox.moe/l7vd0i.png

Anonymous

Anonymous Sun 21 May 2023 04:24:48 No.7386851 Report

Quoted By:

>>7386845
will you be releasing the mix?

Anonymous

Anonymous Sun 21 May 2023 04:31:14 No.7386857 Report

Quoted By: >>7386864

dpm++ 2m sde got added to kdiffusion now. I still can't get a working implementation that actually produces good results and that also respects seeds at different batch sizes, so if somebody else wants to take a look at it please do so.
https://github.com/crowsonkb/k-diffusion/commit/962d62bc73444d005e5828f11ae5b053f287e12e

Anonymous

Anonymous Sun 21 May 2023 04:32:17 No.7386860 Report

Quoted By: >>7386861

is the infinite image browser extension safe? noticing that it has access to my root OS folder by default inside the folder browsing GUI.... >.>

Anonymous

Anonymous Sun 21 May 2023 04:34:15 No.7386861 Report

Quoted By: >>7386885 >>7386921

>>7386860
all extensions are inherently unsafe when you're just pulling from git repos and running unrestricted python code

Anonymous

Anonymous Sun 21 May 2023 04:35:38 No.7386864 Report

Quoted By: >>7386878

>>7386857
nevermind, voldy literally just added it lol. time to see if it actually works because he implemented the exact same way I did which did not work correctly. https://github.com/AUTOMATIC1111/stable-diffusion-webui/commit/31545abe145ac8833c9c15aa41300fb609dcb128

Anonymous

View Same Google ImgOps iqdb SauceNAO 2m v sde v 2m sde.png, 2MiB, 1920x1098

Anonymous Sun 21 May 2023 04:44:52 No.7386878 Report

Quoted By: >>7386887

>>7386864
I dunno, I don't think this is right. it really shouldn't require more sampling steps to look better.

Anonymous

Anonymous Sun 21 May 2023 04:47:29 No.7386882 Report

Quoted By: >>7386919

>>7386821
I need that kirche lora

Anonymous

Anonymous Sun 21 May 2023 04:52:25 No.7386885 Report

Quoted By: >>7386888 >>7386891

>>7386861
such a copout answer. i don't do auto git pulls. i want to know if the app as it's written in the current commit is insecure

Anonymous

Anonymous Sun 21 May 2023 04:53:58 No.7386887 Report

Quoted By: >>7386895

>>7386878
Based on testing here https://gist.github.com/crowsonkb/3ed16fba35c73ece7cf4b9a2095f2b78, for DPM++ 2M SDE, it is suggested to discard the last step, since it fixes some issues.

Anonymous

Anonymous Sun 21 May 2023 04:54:21 No.7386888 Report

Quoted By:

>>7386885
unless you're exposing it to your network then you're probably fine then

Anonymous

Anonymous Sun 21 May 2023 04:56:57 No.7386891 Report

Quoted By: >>7386921

>>7386885
i'm noob as shit, but it seems this file in the repo is doing the heavy lifting: https://github.com/zanllp/sd-webui-infinite-image-browsing/blob/main/scripts/api.py#L44

the access to delete and access files all the way down to root has me scratching my head. maybe there's a way to restrict network access to this extension, but no clue how to target it specifically without knocking out all network access to the a1111 webui

Anonymous

View Same Google ImgOps iqdb SauceNAO 00204-2102462586.png, 489KiB, 504x784

Anonymous Sun 21 May 2023 04:57:27 No.7386892 Report

Quoted By:

Anonymous

Anonymous Sun 21 May 2023 04:58:43 No.7386895 Report

Quoted By: >>7387144

>>7386887
I did that earlier and it helped, I wasn't sure if that's correct though. Kinda sucks the author didn't really leave a comment specifically if that's the "correct" thing to do. They do mention Karras is better though at least, what they mention about the math makes some sense. I'll make a PR just so it gets voldy's attention at least.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00000-3623119755.png, 3MiB, 1280x1792

Anonymous Sun 21 May 2023 05:18:29 No.7386918 Report

Quoted By:

Anonymous

Anonymous Sun 21 May 2023 05:19:09 No.7386919 Report

Quoted By:

>>7386882
still working on it

Anonymous

Anonymous Sun 21 May 2023 05:20:41 No.7386921 Report

Quoted By: >>7386979

>>7386891
if you're running with default settings then you're only running on localhost, so it's not exposed over the network. what you're worried about would be more related to what I said in >>7386861 where you are trusting other extensions not to fuck you over.

Anonymous

Anonymous Sun 21 May 2023 05:42:24 No.7386979 Report

Quoted By: >>7387005

>>7386921
i see, so you're saying another extension could have access to my internet, and use the extensive filesystem access the infinite image browser has to do naughty shit. so there's no sort of UAC style permission that needs to be granted to any of these extensions? i mean i guess not seeing as i didn't grant any permissions or configure anything, yet infinite image browser has access to my whole filesystem

Anonymous

Anonymous Sun 21 May 2023 05:52:07 No.7387005 Report

Quoted By: >>7387030

>>7386979
>and use the extensive filesystem access the infinite image browser has to do naughty shit
any malicious extension can do this, they don't need reliance on the image browser extension, they would implement it themself. it's a separate concern entirely, and why you should be checking extension updates before blindly updating them. network access is more of a concern since then anything outside of SD could potentially do harm, even if the extension is not meant to, but if you're only running on localhost that drops attack vector dramatically.

Anonymous

Anonymous Sun 21 May 2023 06:05:55 No.7387030 Report

Quoted By: >>7387034

>>7387005
so extensions can't get onto the network if a1111 itself is being run locally or did I misunderstand?

Anonymous

View Same Google ImgOps iqdb SauceNAO catbox_gvcy66.png, 1MiB, 1024x1024

Anonymous Sun 21 May 2023 06:08:14 No.7387031 Report

Quoted By:

Anonymous

Anonymous Sun 21 May 2023 06:12:50 No.7387034 Report

Quoted By: >>7387057

>>7387030
extensions can do whatever they want with what is accessible to them on LAN but the attack vector is so low at that point I wouldn't say it's a concern. also a bit dependent on the LAN setup. any other potentially malicious software that is running on localhost wouldn't be trying to target you via SD and instead do their own thing because if they have access to localhost they can already do a lot.
tl;dr check extensions before updating and don't be retarded

Anonymous

View Same Google ImgOps iqdb SauceNAO 1680989934704106.jpg, 300KiB, 1408x1792

Anonymous Sun 21 May 2023 06:12:51 No.7387035 Report

Quoted By: >>7387043

Anonymous

View Same Google ImgOps iqdb SauceNAO 9996490989765848-607199967-killj (...).png, 2MiB, 1600x1600

Anonymous Sun 21 May 2023 06:19:06 No.7387038 Report

Quoted By:

Anonymous

Anonymous Sun 21 May 2023 06:20:20 No.7387040 Report

Quoted By:

I recognize some of you from sdg you fucking degenerates.

Anonymous

Anonymous Sun 21 May 2023 06:23:03 No.7387043 Report

Quoted By: >>7387055

>>7387035
catbox plz

Anonymous

View Same Google ImgOps iqdb SauceNAO 1669147507289673.png, 3MiB, 1280x1792

Anonymous Sun 21 May 2023 06:37:29 No.7387055 Report

Quoted By:

>>7387043
https://files.catbox.moe/1f20yh.png

Anonymous

Anonymous Sun 21 May 2023 06:39:42 No.7387057 Report

Quoted By: >>7387059

>>7387034
sorry for being dense but what you're saying is extensions only have access to LAN and not localhost? if so, seems a bit weird since a1111 itself is running on localhost

Anonymous

Anonymous Sun 21 May 2023 06:41:22 No.7387059 Report

Quoted By:

>>7387057
both. LAN access is dependent on how your network is set up tho.

Anonymous

Anonymous Sun 21 May 2023 07:06:11 No.7387082 Report

Quoted By: >>7387084 >>7387091 >>7387422

Does anyone happen to know of any good Touhou Checkpoints overall? Or at least touhou doujin artist style LORA’s?

Anonymous

Anonymous Sun 21 May 2023 07:07:26 No.7387084 Report

Quoted By: >>7387093

>>7387082
name some doujin artists and/or check gayshit repo in OP

Anonymous

Anonymous Sun 21 May 2023 07:16:19 No.7387091 Report

Quoted By:

>>7387082
There's a Namako (Takorin) one if that's what you're looking for.

Anonymous

Anonymous Sun 21 May 2023 07:16:30 No.7387093 Report

Quoted By: >>7387104

>>7387084
Like Ginyou Haru or Rindou for example.

Anonymous

Anonymous Sun 21 May 2023 07:23:31 No.7387104 Report

Quoted By:

>>7387093
I could've sworn there was one for Rindou, maybe it was under a different name? I haven't seen one for Ginyou Haru/Kamikadou. If you make a dataset someone here could train for you.

Anonymous

Anonymous Sun 21 May 2023 07:38:27 No.7387134 Report

Quoted By:

>>7386589
>>7386590
poorfag cope lmao

Anonymous

Anonymous Sun 21 May 2023 07:45:55 No.7387144 Report

Quoted By: >>7387150

>>7386895
Luckily it got merged. It's good it's on Karras as well, since if you have some lora/model that have a semi "fried" result, without the "discarding" even with Karras it looks artifacted.

Since auto seems to hear you, can you notify him about the wildcard/dynamic prompt issue when using the new hi-res fix?

https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/10565

https://github.com/AUTOMATIC1111/stable-diffusion-webui/commit/ff0e17174f8d93a71fdd5a4a80a4629bbf97f822#commitcomment-114173343

Not sure what's happening, but it seems that the prompt when using the hi-res fix takes prompt of all the variables inside the {} brackets, even if it should just use one.

Anonymous

Anonymous Sun 21 May 2023 07:50:05 No.7387150 Report

Quoted By: >>7387152

>>7387144
I see you opened an issue in the dynamic prompts repo so you probably should wait for an answer there first. I'm not as familiar with how dynamic prompts operates so I don't want to open that can of worms for now.
The hires fix prompt was causing issues with the controlnet extension too but the dev fixed it up pretty quickly after I opened an issue there.

Anonymous

Anonymous Sun 21 May 2023 07:51:45 No.7387152 Report

Quoted By: >>7387153 >>7387156

>>7387150
Sure, gonna see how it goes. Though, the dev of dyanmic prompts seems to be not active since a month ago :( hope he returns and see the issue.

Anonymous

Anonymous Sun 21 May 2023 07:52:56 No.7387153 Report

Quoted By:

>>7387152
2-3 weeks* The last commit was 1 month ago, mb

Anonymous

Anonymous Sun 21 May 2023 07:56:52 No.7387156 Report

Quoted By:

>>7387152
akx is a collaborator on the repo and might see it, he's been doing a shit ton of cleanup on the webui lately

Anonymous

View Same Google ImgOps iqdb SauceNAO file.png, 1MiB, 848x1272

Anonymous Sun 21 May 2023 08:01:34 No.7387159 Report

Quoted By: >>7387160

>>7386787
>ftfy

Anonymous

Anonymous Sun 21 May 2023 08:03:54 No.7387160 Report

Quoted By: >>7387164

>>7387159
>fried that for you

Anonymous

Anonymous Sun 21 May 2023 08:12:38 No.7387164 Report

Quoted By:

>>7387160
KEK
truee

Anonymous

Anonymous Sun 21 May 2023 08:19:18 No.7387170 Report

Quoted By:

>>7386845
dont release the link, I dont want to see more lyco reco, you havent nuked it enough

Anonymous

Anonymous Sun 21 May 2023 08:20:45 No.7387172 Report

Quoted By: >>7387173 >>7387181 >>7387190

Love how when anyone asks for help or needs advice they just get ignored. Lurked tons of these threads and you all circle jerk each other like those faggots from f95.

Anonymous

Anonymous Sun 21 May 2023 08:22:32 No.7387173 Report

Quoted By: >>7388791

>>7387172
Not anyone, anon. Just you <3

Anonymous

View Same Google ImgOps iqdb SauceNAO ysera.png, 1MiB, 768x1152

Anonymous Sun 21 May 2023 08:25:05 No.7387178 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO xyz_grid-0036-3096547508-solo, N (...).jpg, 1MiB, 4000x1653

Anonymous Sun 21 May 2023 08:26:24 No.7387181 Report

Quoted By:

>raising the token limit from 150 to 225 didn't help with the stair thing
well I guess I'll rebake at 0.4 confidence and see if that finally fixes it or somehow or this was RNG all along and never had anything to do with tags.

>>7387172
? what was ignored

Anonymous

Anonymous Sun 21 May 2023 08:37:08 No.7387190 Report

Quoted By: >>7388791

>>7387172
lurk more

Anonymous

Anonymous Sun 21 May 2023 08:41:12 No.7387192 Report

Quoted By: >>7387195 >>7387367

>pull after couple months
>can't gen at the old resolution anymore
i fell for it bros...

Anonymous

Anonymous Sun 21 May 2023 08:42:03 No.7387193 Report

Quoted By:

>he pulled

Anonymous

Anonymous Sun 21 May 2023 08:42:50 No.7387195 Report

Quoted By: >>7387254

>>7387192
I think there's no actual reason for ever pull again for us poorbros

Anonymous

Anonymous Sun 21 May 2023 08:48:43 No.7387201 Report

Quoted By:

sometimes, I pull so hard, I rip the skin

Anonymous

Anonymous Sun 21 May 2023 09:03:08 No.7387212 Report

Quoted By: >>7387234

>pull
>it gets slower
>eats more vram
>I already deleted old extensions so cant roll back
all that's left is to pull more.
Also, do I still need the sd 1.5 ckpt for controlnet? its 7 fucking GB

Anonymous

Anonymous Sun 21 May 2023 09:13:12 No.7387229 Report

Quoted By: >>7387265

It feels much easier to tag images for character lora if you bait yourself to think that you're trying to reverse engineer a prompt for given images, and also omitting character's default look to properly "glue" it to character's name.

Is this common knowledge? I haven't seen it in any Lora training guide.

Anonymous

Anonymous Sun 21 May 2023 09:18:25 No.7387234 Report

Quoted By: >>7387243

>>7387212
What, no? You just need the ControlNet models.
Get pruned fp16 versions, they're way smaller.
Make sure to get the new ControlNet v1.1 ones, you'll find them on HuggingFace.

Anonymous

View Same Google ImgOps iqdb SauceNAO FttI5ByXgAE38tt.jpg, 296KiB, 1280x1536

Anonymous Sun 21 May 2023 09:23:46 No.7387242 Report

Quoted By: >>7387250

Does anyone know which model was used for this?

Anonymous

Anonymous Sun 21 May 2023 09:25:27 No.7387243 Report

Quoted By:

>>7387234
iirc old versions required it or maybe im retarded. time to save some space on this dying drive then

Anonymous

Anonymous Sun 21 May 2023 09:30:21 No.7387250 Report

Quoted By: >>7387291

>>7387242
The fact that the light behind the head has changed color makes me believe that Hassaku

Anonymous

Anonymous Sun 21 May 2023 09:32:36 No.7387252 Report

Quoted By:

>>7386821
Never heard of kirche lora but after seeing this pic will definitely do a sidetrack search.

Anonymous

Anonymous Sun 21 May 2023 09:35:00 No.7387254 Report

Quoted By: >>7387255

>>7387195
Same for richfags desu

Anonymous

Anonymous Sun 21 May 2023 09:35:43 No.7387255 Report

Quoted By:

>>7387254
Controlnet 1.1 seems to be pretty decent

Anonymous

Anonymous Sun 21 May 2023 09:44:50 No.7387264 Report

Quoted By: >>7387265

Does anyone know a Lora/Model to make gigantic tits?

Anonymous

View Same Google ImgOps iqdb SauceNAO xyz_grid-0048-4209822798-solo, N (...).jpg, 337KiB, 4000x377

Anonymous Sun 21 May 2023 09:45:11 No.7387265 Report

Quoted By: >>7387299

Well, I guess this works as a test for batches and generalization aswell - what with people saying batch 1 would overfit on random shit in the background, setting the batch higher doesnt seem to fix this whatsoever.

>>7387229
it's common knowledge that it's a bad practice because it doesn't allow you to change character features as easily.

>>7387264
civitai

Anonymous

Anonymous Sun 21 May 2023 09:48:15 No.7387267 Report

Quoted By: >>7387321

>>7386504
is there like a worklow for this, seems kinda interesting

Anonymous

View Same Google ImgOps iqdb SauceNAO file.png, 2MiB, 1408x1056

Anonymous Sun 21 May 2023 09:55:11 No.7387279 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO 00361-2093643744-nsfw, scan, (tr (...).png, 2MiB, 1024x1536

Anonymous Sun 21 May 2023 09:55:32 No.7387280 Report

Quoted By:

>>7386612
Sorry I was asleep: https://files.catbox.moe/8re578.png
Some image editing and img2img was required.

Anonymous

Anonymous Sun 21 May 2023 09:57:15 No.7387286 Report

Quoted By: >>7387306 >>7387310

Animesharp or remacri? I can't decide myself bros.....

Anonymous

Anonymous Sun 21 May 2023 10:00:12 No.7387291 Report

Quoted By: >>7387298

>>7387250

Can I just say how grateful I am for this community? I ask the same question on Twitter or Discord and get hit with either absolute radio-silence, or people get pissy because they don't want to share their methods, while everyone here seems eager to share their knowledge and discuss their methods with others.

/hdg/ is the Temple in Jerusalem for us coomer-scholars

Anonymous

View Same Google ImgOps iqdb SauceNAO 00023-2816540248.png, 1MiB, 1536x1024

Anonymous Sun 21 May 2023 10:02:32 No.7387298 Report

Quoted By: >>7387315 >>7387372

>>7387291
I present you the Temple of /hgg/

Anonymous

Anonymous Sun 21 May 2023 10:02:54 No.7387299 Report

Quoted By: >>7387309 >>7387320 >>7387322

>>7387265
Both extensive tagging, including with character's default outfit and look, AND omitting default outfit and general look but char's name to "glue" it into are marked as bad practice.

Im not sure what's considered good practice, then.

Anonymous

Anonymous Sun 21 May 2023 10:06:39 No.7387306 Report

Quoted By:

>>7387286
None

Anonymous

Anonymous Sun 21 May 2023 10:07:16 No.7387309 Report

Quoted By: >>7387326 >>7387360

>>7387299
>because it doesn't allow you to change character features as easily

If you do not care about that, you don't have to care about it being "bad practice"

Anonymous

Anonymous Sun 21 May 2023 10:07:43 No.7387310 Report

Quoted By:

>>7387286
Animesharp is decent, but you can always try both and choose for yourself.

Anonymous

Anonymous Sun 21 May 2023 10:10:23 No.7387315 Report

Quoted By:

>>7387298
kek i love it
catbox?

Anonymous

Anonymous Sun 21 May 2023 10:12:04 No.7387320 Report

Quoted By:

>>7387299
>extensive tagging is bad practice
? I think you misunderstood something.
You usually want to tag everything about a character and the outfit, then reduce stray tags such as "skirt, pleated skirt, miniskirt, blue skirt" into the ones that describe it best. An easier case would be with "jewelry, necklace, cross, cross necklace, latin cross" becoming "cross necklace", possibly with "latin cross" added aswell.
And if you mean specific custom outfits then yes they can be reduced to a single custom tag, i.e if a character has a specific school uniform and you don't want it to encroach on generic school uniform tag, then you come up with a custom keyword to reduce it to, much like removing character features to glue the look to the name prompt. Its technically better but often means people will have to remember such keywords when using the lora rather than it being intuitive.

Anonymous

Anonymous Sun 21 May 2023 10:12:07 No.7387321 Report

Quoted By:

>>7387267
controlnet reference only

Anonymous

Anonymous Sun 21 May 2023 10:12:29 No.7387322 Report

Quoted By: >>7387326

>>7387299
Consolidation is good practice to me. Middle way.

Anonymous

Anonymous Sun 21 May 2023 10:17:54 No.7387326 Report

Quoted By: >>7387344 >>7387368

>>7387309
>>7387322
Yeah, feels like trainer is choosing character lora's main purpose. To either produce "wow, that's %character_name%!" in her default look or "wait, is that %character_name%?" while she's wearing torn maid outfit and being shared between faceless (multiple boys).

Anonymous

Anonymous Sun 21 May 2023 10:23:53 No.7387344 Report

Quoted By:

>>7387326
...? Even if you don't remove tags, you can replicate "character name in her default look". That's why removing them is bad practice, as that's the way which locks the other option of giving said character twintails.

Anonymous

Anonymous Sun 21 May 2023 10:28:25 No.7387360 Report

Quoted By:

>>7387309
it really only happens when the lora is fried to shit anyway.. never had a problem changing details with my own shitty loras

Anonymous

Anonymous Sun 21 May 2023 10:32:50 No.7387367 Report

Quoted By: >>7387379

>>7387192
adding --opt-split-attention-v1 fixed my memory problems if any fellow vramlets want to give it a try.

Anonymous

View Same Google ImgOps iqdb SauceNAO 54-545628_hatsune-miku-hatsune-m (...).jpg, 294KiB, 820x973

Anonymous Sun 21 May 2023 10:32:54 No.7387368 Report

Quoted By: >>7387381

>>7387326
Maybe I am misreading what you wrote, but that's not really the purpose of removing tags.

If pic related is the girl you want to train on, then keeping tags like "blue hair, twintails, skirt, etc" will allow you to change these aspects in your prompts later. So you can prompt "red hair" and it will generate her with red hair. Drawback being that you have to prompt "blue hair, twintails, skirt, etc" every time you want the default look.

If you remove tags like "blue hair, twintails, skirt, etc" instead, then the LoRA will always generate that character in her default look, even if you do not prompt "blue hair, twintails, skirt, etc". Biggest drawback being that you will have problems changing everything not tagges, i.e. hair color, outfit, etc.

So you basically choose between "Do I want my LoRA to generate [Attribute] of the character without having to tag [Attribute], risking that I cannot gen the character without [Attribute], or do I want the LoRA to be able to change [Attribute], but therefore having to tag [Attribute] everytime you want it included in the image

Anonymous

View Same Google ImgOps iqdb SauceNAO 00014-1549766813.png, 2MiB, 1536x1024

Anonymous Sun 21 May 2023 10:35:08 No.7387372 Report

Quoted By: >>7387698

>>7387298
Brought together to worship and discuss the power of AI-cooming

Anonymous

Anonymous Sun 21 May 2023 10:36:30 No.7387373 Report

Quoted By:

Using multiple tags is hardly a problem, but you do need to consolidate for quality's sake.

Anonymous

Anonymous Sun 21 May 2023 10:38:35 No.7387379 Report

Quoted By:

>>7387367
>v1
huh

Anonymous

Anonymous Sun 21 May 2023 10:40:25 No.7387381 Report

Quoted By: >>7387395 >>7387421

>>7387368
Its not even that you'll always have to tag it anyways, since it will remember most things naturally, i.e most cases you wont have to prompt "long hair" or "twintails"

Anonymous

Anonymous Sun 21 May 2023 10:48:42 No.7387395 Report

Quoted By: >>7387399 >>7387402 >>7387404

>>7387381
NTA but the question is: will you be able to change it later? If it never learns what long hair is would it then not know how to change it to short hair?

Anonymous

View Same Google ImgOps iqdb SauceNAO kjg5dac6.png, 2MiB, 1024x1536

Anonymous Sun 21 May 2023 10:52:52 No.7387398 Report

Quoted By: >>7387503

Anonymous

Anonymous Sun 21 May 2023 10:54:56 No.7387399 Report

Quoted By:

>>7387395
It can definitely happen.

Anonymous

Anonymous Sun 21 May 2023 10:57:40 No.7387402 Report

Quoted By: >>7387408 >>7387414

>>7387395
NTA but the base model will let you do it.
Regardless, removing or combining any tags (that aren't just incorrect) is ultimately bad training and only for retards that fry their text encoder.

Anonymous

Anonymous Sun 21 May 2023 10:57:57 No.7387404 Report

Quoted By:

>>7387395
im talking about tagging everything, so yes you would have no problem changing it. Just that its an exaggeration that you would need to use every single tag in these cases as it will still mostly associate "char name" with char features

Anonymous

View Same Google ImgOps iqdb SauceNAO 00073-1117785791-masterpiece, be (...).png, 3MiB, 1280x1600

Anonymous Sun 21 May 2023 10:59:07 No.7387406 Report

Quoted By: >>7387419 >>7387424 >>7387812

>enter the forest
>see this
what do?

Anonymous

Anonymous Sun 21 May 2023 10:59:39 No.7387408 Report

Quoted By: >>7387418

>>7387402
>combining tags
thats the first time I see someone speak against consolidating

Anonymous

Anonymous Sun 21 May 2023 11:02:18 No.7387414 Report

Quoted By:

>>7387402
>some guy in the thread tells him he doesn't know what training the text encoder means
>incorporates text encoder into his lexicon as a buzzword for /hdg/ shitposting
lmao

Anonymous

Anonymous Sun 21 May 2023 11:08:03 No.7387418 Report

Quoted By:

>>7387408
It got popular so he's a contrarian about it too now.

Anonymous

Anonymous Sun 21 May 2023 11:09:11 No.7387419 Report

Quoted By:

>>7387406
>exit the forest

Anonymous

Anonymous Sun 21 May 2023 11:10:52 No.7387421 Report

Quoted By:

>>7387381
wont always have to prompt it* derp

Anonymous

Anonymous Sun 21 May 2023 11:11:05 No.7387422 Report

Quoted By:

>>7387082
I don't know what you consider as good, but there's a handful at Civitai. There's a Mega at gayshit that has 2hus like Sukuna and Seija. As for artists, I found anterdel, Baku-P, ke-ta, lolimate, massakasama, and michiking over at gayshit as well. There's probably more, but those are ones that come to mind right now.

Anonymous

Anonymous Sun 21 May 2023 11:11:39 No.7387424 Report

Quoted By:

>>7387406
Run the fuck away
Run the fuck away

Anonymous

Anonymous Sun 21 May 2023 11:12:35 No.7387427 Report

Quoted By: >>7387428 >>7387430 >>7387442 >>7387483

We have Lora training technology for half a year, yet still don't have widely accepted way of training them. I don't believe that there isn't at least one successfull Lora's MEGA/Drive training folder with dataset to serve a good example of how to tag shit instead of referring to various training guides with contradicting information in them.

Anonymous

Anonymous Sun 21 May 2023 11:13:45 No.7387428 Report

Quoted By:

>>7387427
>blackbox ml
>accepted way
lmao

Anonymous

Anonymous Sun 21 May 2023 11:14:28 No.7387430 Report

Quoted By:

>>7387427
Just trial and error shit yourself, everyone else could be trolling or retarded so you may as well assume they are both

Anonymous

Anonymous Sun 21 May 2023 11:25:54 No.7387440 Report

Quoted By: >>7387461

>>7386754
Link? If it's actually a better base model than NAI then it could be huge; though I have little hope.

Anonymous

Anonymous Sun 21 May 2023 11:28:39 No.7387442 Report

Quoted By: >>7387476 >>7387513

>>7387427
Lora training is a black box, and there's no objective way to evaluate improvements from tagging across all datasets due to so many variables that can be tweaked for training.

Even adding a single image can drastically fuck your lora up sometimes. I once baked a character lora where adding a particular high quality image from official art would for some reason, fuck up the lora character's eyes. Removing it from the dataset fixed the issue.

Point is, lora training is so black box you can't tell what works if the improvements aren't completely consistent+the lora is already decent to begin with.

Just use whichever method you like and want to spend more/less time on.

Anonymous

Anonymous Sun 21 May 2023 11:29:27 No.7387443 Report

Quoted By: >>7387445 >>7387449

Clearly we need our furfags saviors to teach us model training.
So over, animebros...

Anonymous

Anonymous Sun 21 May 2023 11:30:02 No.7387445 Report

Quoted By: >>7387447

>>7387443
you could at least try typing differently

Anonymous

Anonymous Sun 21 May 2023 11:33:18 No.7387447 Report

Quoted By:

>>7387445
But it's so over...

Anonymous

View Same Google ImgOps iqdb SauceNAO 00010-3481400946_cleanup.png, 2MiB, 1024x1536

Anonymous Sun 21 May 2023 11:37:39 No.7387449 Report

Quoted By: >>7387456 >>7387476

>>7387443
Tried looking at furfags general at trash and couldn't handle seeing that disgusting shit. Whatever they're doing, they and (You) can fuck off.
Checked vtai though and they do have some decent mixes/finetunes for 1girl posting. Easier than having to deal with 9k loras for each character that's for sure.

Anonymous

Anonymous Sun 21 May 2023 11:42:27 No.7387456 Report

Quoted By: >>7387470 >>7387485

>>7387449
People can't get obvious jokes, huh?

Anonymous

Anonymous Sun 21 May 2023 11:44:01 No.7387461 Report

Quoted By:

>>7387440
i skimmed the archives and it looks like i misread something

Anonymous

Anonymous Sun 21 May 2023 11:49:43 No.7387470 Report

Quoted By: >>7387473

>>7387456
>I was only pretending to be retarded!!!

Anonymous

View Same Google ImgOps iqdb SauceNAO 00005-410368326.png, 2MiB, 1024x1536

Anonymous Sun 21 May 2023 11:51:57 No.7387472 Report

Quoted By: >>7387490 >>7387538

thinking of doing a hentai card based game, kinda like Magic The Gathering, using SD.

I have quite a few already

Anonymous

Anonymous Sun 21 May 2023 11:53:02 No.7387473 Report

Quoted By:

>>7387470
i mean yeah it was pretty obvious lmao
are you stupid

Anonymous

View Same Google ImgOps iqdb SauceNAO 00003-3114407537.png, 2MiB, 1024x1536

Anonymous Sun 21 May 2023 11:53:10 No.7387475 Report

Quoted By:

Anonymous

Anonymous Sun 21 May 2023 11:53:37 No.7387476 Report

Quoted By: >>7387484 >>7387487

>>7387442
I felt that when my mistagged 25 images lora with 10 and 20 repeats performed much better than 427 images one with 1-5 repeats, but with 2x repeats overall.

Yet still I cannot blame software or call it purely a blackbox because I don't have good knowledge of how to use it.

>>7387449
Pickle chan...

There are some good anime-style anhtro arts in /trash/ (with Kagemori Michiru raccoon girl, to be exact), but you have to put yourself through overcooked blurry muscle pokemons stuffing theirselves with horse cocks, close-up male dragon anuses, photorealistic bestiality, people fucking overly detailed, thick and curvy scp-1471 and other shit.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00056-3982046907.png, 2MiB, 1024x1536

Anonymous Sun 21 May 2023 11:53:44 No.7387477 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO 00060-3949533320.png, 2MiB, 1024x1536

Anonymous Sun 21 May 2023 11:54:18 No.7387479 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO 00061-1339792501.png, 2MiB, 1024x1536

Anonymous Sun 21 May 2023 11:54:51 No.7387481 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO 00078-1741287092.png, 2MiB, 1024x1536

Anonymous Sun 21 May 2023 11:55:26 No.7387482 Report

Quoted By:

Anonymous

Anonymous Sun 21 May 2023 11:55:42 No.7387483 Report

Quoted By: >>7387495

>>7387427
well, alot of people insist on various dumb shit without testing anything, like the batch size debates.
As for tagging most of it is known but "overtagging" might be an issue people slept on - since imma be re-doing the training on higher confidence later tonight for the weird "lora refuses to generate stairs" issue I think we can get somewhat closer to some truth if high confidence actually improves it and it wasnt just a lucky RNG bake.

Anonymous

Anonymous Sun 21 May 2023 11:56:19 No.7387484 Report

Quoted By:

>>7387476
The contrast of some of the best and most creative images and then a blurry 144p Machoke taking a horse cock in the ass is pretty funny.

Anonymous

Anonymous Sun 21 May 2023 11:56:47 No.7387485 Report

Quoted By: >>7387489

>>7387456
Maybe that anon is on the spec

Anonymous

View Same Google ImgOps iqdb SauceNAO 00116-1339453159.png, 2MiB, 1024x1536

Anonymous Sun 21 May 2023 11:58:03 No.7387486 Report

Quoted By:

Anonymous

Anonymous Sun 21 May 2023 11:58:28 No.7387487 Report

Quoted By: >>7387495 >>7387496

>>7387476
epochs vs repeats influence the outcome, its not necessarily less repeats and more epochs = better in every case, but more epochs is more likely to give you a good result.
similarly im starting to think more pictures are not actually going to result in an improvement, my best loras were small ones

Anonymous

Anonymous Sun 21 May 2023 11:59:29 No.7387489 Report

Quoted By: >>7387507

>>7387485
Or trying to fit in.

Anonymous

Anonymous Sun 21 May 2023 12:00:33 No.7387490 Report

Quoted By: >>7387492 >>7387495 >>7387832

>>7387472
A few years ago I used to think that art and music will be the reasons why I won't be able to make a game on my own. And now I'm capable of making the music I want, I have an infinite art generator and the only thing I'm getting filtered by is programming

Anonymous

View Same Google ImgOps iqdb SauceNAO 00111-197237215.png, 2MiB, 1024x1536

Anonymous Sun 21 May 2023 12:01:20 No.7387492 Report

Quoted By:

>>7387490
luckily that is the one thing I can do

Anonymous

Anonymous Sun 21 May 2023 12:03:12 No.7387495 Report

Quoted By: >>7387497 >>7387498 >>7387499 >>7387502

>>7387483
>like the batch size debates
Isn't it only increases training speed for the sake of using more VRAM?

>>7387487
Both were trained up to e10.

>>7387490
Don't tell him about ChatGPT.

Anonymous

Anonymous Sun 21 May 2023 12:06:18 No.7387496 Report

Quoted By:

>>7387487
I don't go over 150 images myself.

Anonymous

Anonymous Sun 21 May 2023 12:06:20 No.7387497 Report

Quoted By: >>7387500

>>7387495
>Isn't it only increases training speed for the sake of using more VRAM?
No. Results degrade as you increase the batch size.

Anonymous

Anonymous Sun 21 May 2023 12:06:52 No.7387498 Report

Quoted By:

>>7387495
No. Batches in training have the AI process more images then add the average of their weights to the model you are making. Obviously changing the number of images you are averaging will change the result.

Anonymous

Anonymous Sun 21 May 2023 12:07:04 No.7387499 Report

Quoted By: >>7388280 >>7388292

>>7387495
ChatGPT works very well if you already know how to code, can break down complex concepts into simple discrete tasks to prompt separately, and know what you're looking for.

Otherwise you're going to have a bad time with all the weird ass non-existent libraries ChatGPT dreams up, among other LLM quirks.

We'll need to give it a few more years.

Anonymous

Anonymous Sun 21 May 2023 12:07:23 No.7387500 Report

Quoted By: >>7387504

>>7387497
Except batch 1 which looks bad most of the time

Anonymous

Anonymous Sun 21 May 2023 12:08:23 No.7387502 Report

Quoted By: >>7387516

>>7387495
>Both were trained up to e10.
as in epoch 10? you said one had more steps than the other. Well, not like it matters anyway - its just that epochs shuffle every new one, so certain things would get "baked" harder with more repeats compared to epochs. It's a bad thing more often that it isnt (RNG), but sometimes it could have had a positive outcome

Anonymous

Anonymous Sun 21 May 2023 12:09:23 No.7387503 Report

Quoted By: >>7388265

>>7387398
catbox?

Anonymous

Anonymous Sun 21 May 2023 12:09:25 No.7387504 Report

Quoted By: >>7387523

>>7387500
prove it or go shitpost elsewhere tonight

Anonymous

Anonymous Sun 21 May 2023 12:10:46 No.7387507 Report

Quoted By: >>7387508 >>7387510

>>7387489
Or just tired of unfunny faggot shilling furry garbage and playing it off as a joke when people get annoyed.

Anonymous

Anonymous Sun 21 May 2023 12:11:30 No.7387508 Report

Quoted By: >>7387512

>>7387507
NTA but it was obviously a sarcastic joke you autismo, get over it.

Anonymous

Anonymous Sun 21 May 2023 12:11:52 No.7387510 Report

Quoted By:

>>7387507
It couldn't been more obvious shitposting if I tried, sperg.
Actual spec.

Anonymous

Anonymous Sun 21 May 2023 12:12:07 No.7387512 Report

Quoted By:

>>7387508
Wow, now I totally get it. Laughed my fucking ass off.

Anonymous

Anonymous Sun 21 May 2023 12:12:34 No.7387513 Report

Quoted By: >>7387518

>>7387442
>Removing it from the dataset fixed the issue

Are you sure it's not just the lora being trained on a different seed?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00112-3846418261.png, 2MiB, 1024x1536

Anonymous Sun 21 May 2023 12:12:47 No.7387514 Report

Quoted By:

Anonymous

Anonymous Sun 21 May 2023 12:15:58 No.7387516 Report

Quoted By: >>7387524

>>7387502
>certain things would get "baked" harder with more repeats compared to epochs. It's a bad thing more often that it isnt
Funny to hear that as I'm training a 72 images Lora with same 10 and 20 repeats.
It will be crispy as fuck I guess.

Anonymous

Anonymous Sun 21 May 2023 12:16:59 No.7387518 Report

Quoted By:

>>7387513
NTA but prob not, I've had it across a few seeds because official art clashed with how fanart drew eyes resulting in blurred mess since the AI was stuck halfway between official eye and fanart eye.
Its a pretty obvious thing to be wary of.

Anonymous

Anonymous Sun 21 May 2023 12:18:46 No.7387520 Report

Quoted By:

I should actually try 1 batch again to see whether it helps with asymmetric features.

Anonymous

Anonymous Sun 21 May 2023 12:23:44 No.7387523 Report

Quoted By: >>7387535

>>7387504
i'm not shitposting, i used to be a vramlet so i did many many batch 1 bakes, and then i upgraded and did higher batch sizes and i see way less overfitting on small elements of single pictures. however i have settled on some different training settings over the last 2 months, so i think it's worth testing again

Anonymous

Anonymous Sun 21 May 2023 12:24:12 No.7387524 Report

Quoted By:

>>7387516
No, its nothing that obvious. Its harder to control these loras as it does a massive leap from one epoch to another so it might get overbaked in that sense, since it can jump from underbaked to overbaked.
Peple xplained this way back, but essentially it runs in the same order until an epoch reshuffles it - the order in which it runs through images affects how it learns. A certain order can either make it stuck on a mistake, like some hair flower being learned wrong, or it can be beneficial and make it learn some detail better. More epochs means more shuffles, so naturally that *should* lead to it avoiding these pitfalls and eventually learning better, even if it *can* end up a bad thing.
At the end of the day it doesnt really matter that much - moreso its just better to have more epochs more often as each will look different allowing you to pick the best one

Anonymous

Anonymous Sun 21 May 2023 12:28:44 No.7387528 Report

Quoted By: >>7387542

Also worth saying that repeats have almost no practical use besides weighing different folders, and are the equivalent of duplicating the dataset.
Imagine the shitting if someone posted a dataset with 10x the same pictures, but then it's okay to do to with repeats.

Anonymous

View Same Google ImgOps iqdb SauceNAO xyz_grid-0044-3706654609-solo, N (...).jpg, 306KiB, 4000x377

Anonymous Sun 21 May 2023 12:34:53 No.7387535 Report

Quoted By: >>7387537 >>7387539

>>7387523
>overfitting on small elements
that sounds like bullshit that has nothing to do with batches, at least it doesnt really seem to matter as far as fixing overfitting goes. See pic rel, im pretty sure like 2 images in the dataset are causing "sitting" to overfit into a particular chair or wooden floor, if what you claim is true then raising batches should have fixed it or remedied it even a little bit, but it runs into the exact same issue with the exact same severity.

Anonymous

Anonymous Sun 21 May 2023 12:36:17 No.7387537 Report

Quoted By:

>>7387535
yeah i think it might actually have more to do with tagging and tagging confidence etc than anything else, like i said, i've learned a lot recently so i think it's worth revisiting

Anonymous

View Same Google ImgOps iqdb SauceNAO 1675352378362292.png, 4MiB, 1408x1792

Anonymous Sun 21 May 2023 12:36:47 No.7387538 Report

Quoted By:

>>7387472
What's the catbox for the card format, looks fun.

Anonymous

Anonymous Sun 21 May 2023 12:37:34 No.7387539 Report

Quoted By: >>7387542

>>7387535
It's all about normalization with batches. But it's hard to say whether it's gonna be bad or not like with every parameter.

Anonymous

Anonymous Sun 21 May 2023 12:41:08 No.7387542 Report

Quoted By: >>7387549

>>7387528
epochs are duplicating the dataset too though, silly.

>>7387539
Yeah but it should have "normalized"/"generalized" the sitting tag across several sitting images in the dataset - its not all with the same chair.
While I don't deny it generalizes, you have to question how high would the batch size have to be to do it properly so as to fix such issues - and at that point, would it not generalize so much that the entire training is fucked?

Anonymous

Anonymous Sun 21 May 2023 12:44:14 No.7387549 Report

Quoted By: >>7387552

>>7387542
>epochs are duplicating the dataset too though, silly.
Except they literally don't. They extend training, but the image amount per epoch is the same.
That's why repeats are deleterious. You can get the same image x times in a row per shuffle and that's what overfits it.
Epochs just shuffle the same singular images more times.
>>7387542
It depends on dataset size I think. Higher batches fuck up low image datasets.
IMO the less batches the better. Logically and practically.

Anonymous

Anonymous Sun 21 May 2023 12:47:50 No.7387552 Report

Quoted By: >>7387554

>>7387549
? If you do 1 repeat on 2 epochs, you will have ran through every image twice, essentially duplicating it.
If you do 2 repeats on 1 epoch, you will run through every image twice, essentially duplicating it.
The only thing that changed is that they were shuffled once. We do not know which of these will result in image X being ran more times in a row as thats just RNG since each epoch can in theory shuffle in the worst way possible.

Anonymous

Anonymous Sun 21 May 2023 12:51:26 No.7387554 Report

Quoted By: >>7387558 >>7387563

>>7387552
But we do. You misunderstand how it works.
It shuffles the images multiplied per repeats each epoch.
If you have 10 images you get a shuffle of 10 images
If you have 10 images times 10 repeats, it means you can get up to 10 of the same images in a row per shuffle which is per epoch.
The worst scenario in 1 repeat and multiple epochs is getting the same image at the end and start on an epoch.

Anonymous

Anonymous Sun 21 May 2023 12:54:46 No.7387558 Report

Quoted By:

>>7387554
damn, I went full retard. lol True, it cant chain the same image more than twice in a row.

Anonymous

Anonymous Sun 21 May 2023 13:00:27 No.7387563 Report

Quoted By: >>7387569 >>7387576

>>7387554
So I could went with 5 and 10 repeats but 20 epoches instead to be able to choose between twice more epoches with less apparent drawbacks than possibly overfitting with 2x more repeats.

Anonymous

Anonymous Sun 21 May 2023 13:02:29 No.7387569 Report

Quoted By:

>>7387563
Basically. I use 1 as a conclusion to it.

Anonymous

View Same Google ImgOps iqdb SauceNAO Rikku X-2 v3 - locon-sv_ratio-10 (...).png, 1MiB, 768x1152

Anonymous Sun 21 May 2023 13:06:37 No.7387572 Report

Quoted By: >>7387573

Uploaded Rikku

https://mega.nz/folder/WjwmjSpA#_Sr_Gpl7WBXwCeSvV0eEtw

Anonymous

Anonymous Sun 21 May 2023 13:07:34 No.7387573 Report

Quoted By:

>>7387572
cute al bhed

Anonymous

View Same Google ImgOps iqdb SauceNAO 00077-1292121624.png, 572KiB, 512x768

Anonymous Sun 21 May 2023 13:09:44 No.7387575 Report

Quoted By: >>7387579 >>7387581

How do I prevent the characters face from being in dark shadows in every image?

Anonymous

Anonymous Sun 21 May 2023 13:10:33 No.7387576 Report

Quoted By:

>>7387563
I settled on 6.

Anonymous

Anonymous Sun 21 May 2023 13:12:18 No.7387579 Report

Quoted By: >>7387582

>>7387575
Try monochrome, greyscale in negatives.
Also might be your model or lora.

Anonymous

Anonymous Sun 21 May 2023 13:15:01 No.7387581 Report

Quoted By: >>7387582

>>7387575
Unironically try prompting a dark background

Anonymous

Anonymous Sun 21 May 2023 13:15:58 No.7387582 Report

Quoted By:

>>7387579
>>7387581

It seems to be the prompt "day" for some reason...

Anonymous

View Same Google ImgOps iqdb SauceNAO 00023-2347496995ed.png, 813KiB, 1024x1024

Anonymous Sun 21 May 2023 13:20:40 No.7387585 Report

Quoted By: >>7387586 >>7387588

Anonymous

Anonymous Sun 21 May 2023 13:21:42 No.7387586 Report

Quoted By: >>7387590

>>7387585
gato

Anonymous

View Same Google ImgOps iqdb SauceNAO ex.jpg, 919KiB, 2426x1536

Anonymous Sun 21 May 2023 13:24:09 No.7387587 Report

Quoted By: >>7387713

Can anyone help me achieve the style of the image on the right? Left is the best I could achieve. I'm pretty sure "hassakuHentaiModel" is the right model, though I'm wondering if they used a different VAE, or if I'm missing something in the prompt

https://files.catbox.moe/r4mzv3.png

Anonymous

Anonymous Sun 21 May 2023 13:24:39 No.7387588 Report

Quoted By: >>7387590

>>7387585
nice lycoris recoil lora

Anonymous

Anonymous Sun 21 May 2023 13:26:03 No.7387590 Report

Quoted By: >>7387592

>>7387586
neetbox
https://files.catbox.moe/lc1oaq.png
https://files.catbox.moe/d226y0.png
>>7387588
It doesn't even look like lycoris

Anonymous

Anonymous Sun 21 May 2023 13:28:13 No.7387592 Report

Quoted By: >>7387593 >>7387594

>>7387590
wdym, thats Takina from lycoric recoil with a few details changed

Anonymous

Anonymous Sun 21 May 2023 13:29:29 No.7387593 Report

Quoted By:

>>7387592
that's actually meg griffin so you're mistaken

Anonymous

Anonymous Sun 21 May 2023 13:29:39 No.7387594 Report

Quoted By:

>>7387592
>thats [black hair girl with bangs] from [A1 Pictures show]

Anonymous

View Same Google ImgOps iqdb SauceNAO 00022-1041456382.png, 2MiB, 1024x1536

Anonymous Sun 21 May 2023 13:30:17 No.7387595 Report

Quoted By: >>7387596

frog

Anonymous

Anonymous Sun 21 May 2023 13:30:35 No.7387596 Report

Quoted By:

>>7387595
frogposter

Anonymous

View Same Google ImgOps iqdb SauceNAO 00042-1823817481.png, 2MiB, 1024x1536

Anonymous Sun 21 May 2023 13:31:17 No.7387597 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO 00209-3492882040.png, 2MiB, 1152x1664

Anonymous Sun 21 May 2023 13:35:52 No.7387600 Report

Quoted By: >>7387609 >>7387610 >>7387619 >>7387622 >>7387961

>https://mega.nz/folder/ZvA21I7L#ZZzU42rdAyWFOWQ_O94JaQ/folder/0io1WQTa
Baked a catgirl lingerie lora because base models are a little inconsistent and outright can't do the cat ear flaps on the panties.
Yes. I'm aware you can prompt catgirl keyhole bras easily without using loras. Just not the panties, which is also what I'm after.

Prompting colours other than black and white will be difficult due to dataset bias. Bonus occasional cat ear thighhighs, which was completely unintended but seems to be a thing the lora picked up on anyway.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00024-399561070.png, 2MiB, 1024x1536

Anonymous Sun 21 May 2023 13:38:00 No.7387603 Report

Quoted By: >>7387606 >>7387616 >>7387649

Your crippling obesity has progressed to the point that are no longer able to wash yourself.

Luckily for you, your neetbucks enable you to pay a poor girl to do the job for you. Today's her first day

Anonymous

View Same Google ImgOps iqdb SauceNAO 00162.png, 2MiB, 1120x1680

Anonymous Sun 21 May 2023 13:41:30 No.7387605 Report

Quoted By: >>7388162

Anonymous

Anonymous Sun 21 May 2023 13:41:34 No.7387606 Report

Quoted By:

>>7387603
based

Anonymous

Anonymous Sun 21 May 2023 13:43:10 No.7387609 Report

Quoted By:

>>7387600
nice, i noticed that it codulnt' do that panties myself

Anonymous

Anonymous Sun 21 May 2023 13:43:41 No.7387610 Report

Quoted By:

>>7387600
Extremely based, thanks.

Anonymous

Anonymous Sun 21 May 2023 13:46:48 No.7387616 Report

Quoted By:

>>7387603
>american porn

Anonymous

View Same Google ImgOps iqdb SauceNAO 00053-2853706609.png, 1MiB, 896x1344

Anonymous Sun 21 May 2023 13:48:25 No.7387619 Report

Quoted By: >>7388162

>>7387600
thank you anon

Anonymous

View Same Google ImgOps iqdb SauceNAO catbox_f72gog.png, 1MiB, 1024x1280

Anonymous Sun 21 May 2023 13:48:52 No.7387620 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO 00230-1278963287.png, 2MiB, 1152x1664

Anonymous Sun 21 May 2023 13:49:35 No.7387622 Report

Quoted By: >>7387781

>>7387600
Also baked a panties aside lora.

But I'm going to do more testing first to figure out what I can/can't prompt with it, since it seems to only work with pov missionary at the moment.

Anonymous

View Same Google ImgOps iqdb SauceNAO A6C41FA9-6B7E-4A55-9371-D52EBB17 (...).jpg, 362KiB, 1773x1773

Anonymous Sun 21 May 2023 13:51:21 No.7387624 Report

Quoted By: >>7387629 >>7387650

Which models are mainly used these days? I’ve been using the never ending dream model, but it’s not really a hentai model.

Anonymous

View Same Google ImgOps iqdb SauceNAO 02979-2618028943.png, 2MiB, 1024x1536

Anonymous Sun 21 May 2023 13:52:50 No.7387627 Report

Quoted By: >>7387630 >>7387648

somebody help, I can't stop proompting

my house is a literal pigsty, I haven't showered in more than a week

Anonymous

Anonymous Sun 21 May 2023 13:52:52 No.7387628 Report

Quoted By: >>7387636 >>7387637

https://danbooru.donmai.us/wiki_pages/concentration_camp

>only 3 images

bummer

Anonymous

Anonymous Sun 21 May 2023 13:53:52 No.7387629 Report

Quoted By:

>>7387624
don't upscale like that senpai, it smears everything, use hires fix

I've been loving the based65 model, personally

Anonymous

Anonymous Sun 21 May 2023 13:54:14 No.7387630 Report

Quoted By: >>7387638 >>7387648

>>7387627
it looks kinda vaeless
>pigsty
you get used to it

Anonymous

View Same Google ImgOps iqdb SauceNAO 00101-66468004.png, 2MiB, 1024x1536

Anonymous Sun 21 May 2023 13:54:51 No.7387634 Report

Quoted By: >>7387640

>fail for hours to get a properly detailed bar-like background
>later prompt masterpiece, best quality, 1girl, (in jar:1.4) just for fun and receive this

the ai is trolling me

Anonymous

Anonymous Sun 21 May 2023 13:55:15 No.7387636 Report

Quoted By:

>>7387628
check pixiv? controlnet irl images with anime lineart? (lol)

Anonymous

Anonymous Sun 21 May 2023 13:56:24 No.7387637 Report

Quoted By:

>>7387628
you can invoke stable diffusion's concepts most likely

Anonymous

Anonymous Sun 21 May 2023 13:56:46 No.7387638 Report

Quoted By: >>7387644 >>7387648

>>7387630
I strongly prefer the pale colors of mild vaes like the anything v4 I use, compared to the overcooked colors that seem very popular on here

Anonymous

Anonymous Sun 21 May 2023 13:57:51 No.7387640 Report

Quoted By: >>7387653

>>7387634
"Customer, I tried to open the jar and it broke my fingers!"

I believe lucky hits like that happen because the seed is of a very similar image, like 90% similarity.

Anonymous

Anonymous Sun 21 May 2023 13:59:32 No.7387644 Report

Quoted By: >>7387648

>>7387638
I guess the gamma of one's screen also makes a lot of difference

Anonymous

View Same Google ImgOps iqdb SauceNAO 02981-2618028943.png, 2MiB, 1024x1536

Anonymous Sun 21 May 2023 14:01:22 No.7387648 Report

Quoted By: >>7387655 >>7387666

>>7387627
>>7387630
>>7387638
>>7387644
same exact image but with the "mpaffl perfect color" vae, it clearly looks overcooked to me, is it my screen?

Anonymous

Anonymous Sun 21 May 2023 14:01:28 No.7387649 Report

Quoted By:

>>7387603
The Whale

Anonymous

Anonymous Sun 21 May 2023 14:01:42 No.7387650 Report

Quoted By: >>7387651

>>7387624
our models are obsolete compared to furrychads, it's over

Anonymous

Anonymous Sun 21 May 2023 14:02:29 No.7387651 Report

Quoted By:

>>7387650
save us furrygods pleaaaseee

Anonymous

Anonymous Sun 21 May 2023 14:02:51 No.7387652 Report

Quoted By:

furrytrash models are nothing compared to /tv/'s sneedposting finetunes. it truly is over

Anonymous

View Same Google ImgOps iqdb SauceNAO 00107-2446627557.png, 516KiB, 512x768

Anonymous Sun 21 May 2023 14:02:58 No.7387653 Report

Quoted By: >>7387660

>>7387640
nah, it regularly spits out more detailed bar-like backgrounds with just this prompt

Anonymous

View Same Google ImgOps iqdb SauceNAO 00111-4070268130.png, 1MiB, 1536x1024

Anonymous Sun 21 May 2023 14:03:44 No.7387654 Report

Quoted By:

Anonymous

Anonymous Sun 21 May 2023 14:03:58 No.7387655 Report

Quoted By: >>7387681

>>7387648
it removed the shitty blur, it looks fine. Adjust your screen and see for yourself

Anonymous

View Same Google ImgOps iqdb SauceNAO 00106-3102083336.png, 548KiB, 512x768

Anonymous Sun 21 May 2023 14:09:03 No.7387660 Report

Quoted By:

>>7387653
With a style LoRA it produces pretty cool looking jar's too

Anonymous

Anonymous Sun 21 May 2023 14:14:14 No.7387666 Report

Quoted By: >>7387681

>>7387648
It is, but at the same time it removes noise effect and blur. Try playing with (flat color) a bit to get less toxic colors.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00139-1569140448.png, 2MiB, 1024x1536

Anonymous Sun 21 May 2023 14:23:29 No.7387677 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO dehaze comparison.png, 901KiB, 1024x602

Anonymous Sun 21 May 2023 14:27:00 No.7387681 Report

Quoted By: >>7387688 >>7387689

>>7387655
>>7387666

it seems to me that the clear winner is the camera raw filter dehaze from photoshop, the perfect color vae clearly overcooks

I'll have to remember to give every pic I make a quick pass on the photoshop dehaze

Anonymous

Anonymous Sun 21 May 2023 14:28:39 No.7387683 Report

Quoted By: >>7387692 >>7387696 >>7387697 >>7387699

>check anon's catbox
>prompt filled with placebo
>only three actual danbooru tags
>hard carried by character lora

Anonymous

Anonymous Sun 21 May 2023 14:29:50 No.7387688 Report

Quoted By: >>7387700 >>7387710

>>7387681
photoshop is the worst out of all 3 actually. It has the worst parts of both, looks artificial and you would be insane to actually bother with this for every picture.

Anonymous

Anonymous Sun 21 May 2023 14:31:00 No.7387689 Report

Quoted By: >>7387700

>>7387681
you know theres more vae than just these two right?
maybe monochrome at super low strength could help

Anonymous

Anonymous Sun 21 May 2023 14:32:19 No.7387691 Report

Quoted By: >>7387701

Is there a vae that doesn't change details that much and just only fix purple spots

Anonymous

Anonymous Sun 21 May 2023 14:32:35 No.7387692 Report

Quoted By: >>7387695

>>7387683
I've jumped back on forth on quality tags. Sometimes it does seem like they do improve things, but other times it looks like they make things so much worse.

Anonymous

Anonymous Sun 21 May 2023 14:35:00 No.7387695 Report

Quoted By:

>>7387692
quality tags I'm iffy on, but occasionally I still see (perfect face:1.2) which I can't believe does anything

Anonymous

Anonymous Sun 21 May 2023 14:35:20 No.7387696 Report

Quoted By:

>>7387683
If they like the results of it - why bother?

Anonymous

Anonymous Sun 21 May 2023 14:36:32 No.7387697 Report

Quoted By:

>>7387683
Reminds me one time I asked for a catbox of a good gen only to find out its a mess of prompts. I'm actually impressed they managed to get a good gen out of it.

Anonymous

Anonymous Sun 21 May 2023 14:36:35 No.7387698 Report

Quoted By: >>7387704

>>7387372
prompt for the crowd?

Anonymous

Anonymous Sun 21 May 2023 14:36:42 No.7387699 Report

Quoted By: >>7387706

>>7387683
do tags only work if they match exact booru tags? i have the tag autofiller extension and sometimes what i type will vary slightly from what it suggests

Anonymous

Anonymous Sun 21 May 2023 14:37:14 No.7387700 Report

Quoted By:

>>7387688
...you guys are literally impossible to satisfy

Haze is used in art, it has a purpose, it doesn't look bad if it blends in with the dreamy context of an artwork, as long as it doesn't make it illegible.

How is middle the worst when it's literally just left with less gray overlay? Right is literally burnt. Look at the hair in-between the eyes and the mouth, or how the single bottom eyelashes stand out isolated

>>7387689
Those are the big changes among the vaes I use, they either make it pale, overcooked, or introduce the dreaded purple glitches

Anonymous

Anonymous Sun 21 May 2023 14:40:54 No.7387701 Report

Quoted By: >>7387707 >>7387708 >>7387882

>>7387691
vaes don't change details, the little details change you see when reprompting are caused by the gacha nature of stable diffusion. I think that disabling xformers makes it more constant when reprompting with the same seed, but way slower.

as for the purple splotches I use:
-anything v4 (pale colors)
-animevae (stolen from novelai)
-mse 840000 for 3dpd
-mpaffl perfect color for bright colors but be careful not to overcook

Anonymous

Anonymous Sun 21 May 2023 14:44:00 No.7387704 Report

Quoted By:

>>7387698
literally just crowd, crowded

https://files.catbox.moe/4z4itf.png

Anonymous

Anonymous Sun 21 May 2023 14:44:07 No.7387706 Report

Quoted By:

>>7387699
for example, if you wanna prompt erect nipples under clothes - (covered nipples) usually produces best results, and it's a danbooru tag

Anonymous

Anonymous Sun 21 May 2023 14:45:09 No.7387707 Report

Quoted By: >>7387708

>>7387701

>anything v4

Is that the anything-v4.0.ckpt vae from huggingface?

Anonymous

Anonymous Sun 21 May 2023 14:46:21 No.7387708 Report

Quoted By:

>>7387701
>>7387707
I meant the anything-v4.0.vae.pt vae

Anonymous

Anonymous Sun 21 May 2023 14:47:26 No.7387710 Report

Quoted By:

>>7387688
it just feels less vibrant than vae version, you would think that anon already slapped vae on second pic if he didn't show perfect color vae version

Anonymous

Anonymous Sun 21 May 2023 14:49:56 No.7387712 Report

Quoted By:

imminent_ tags are probably the hardest to do. I've had no luck with control nets or anything.

Anonymous

Anonymous Sun 21 May 2023 14:51:13 No.7387713 Report

Quoted By: >>7387720

>>7387587
anyone?

Anonymous

Anonymous Sun 21 May 2023 15:00:57 No.7387720 Report

Quoted By: >>7387722

>>7387713
It reminds me of this style I used once, but its not this.
https://civitai.com/models/38459/style

Anonymous

Anonymous Sun 21 May 2023 15:03:20 No.7387722 Report

Quoted By:

>>7387720
I'm pretty sure they didn't use a LoRA, as they are genning on a Discord I think. That's why I was thinking it's probably prompt or vae related

Anonymous

Anonymous Sun 21 May 2023 15:07:50 No.7387726 Report

Quoted By:

>>7386484
all of mine seem to come out super dark for some reason

Anonymous

Anonymous Sun 21 May 2023 15:24:04 No.7387742 Report

Quoted By: >>7387747

Where do I find R-ESRGAN 4x+ Anime6B upscale model? Doesn't look like I have it in auto1111.

This one?

https://github.com/xinntao/Real-ESRGAN/blob/master/docs/anime_model.md

Anonymous

Anonymous Sun 21 May 2023 15:27:59 No.7387747 Report

Quoted By: >>7387762

>>7387742
In settings > upscaling there's a bunch of R-ESRGAN upscalers you have to enable and then restart a1111 for them to show up in the dropdown

Anonymous

View Same Google ImgOps iqdb SauceNAO 00857-2596216751.png, 1MiB, 1024x1536

Anonymous Sun 21 May 2023 15:30:31 No.7387753 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO 00870-2319887860.png, 2MiB, 1296x1536

Anonymous Sun 21 May 2023 15:31:35 No.7387757 Report

Quoted By: >>7387767 >>7390006

This took a ridiculous amount of time to do. I tried tile diffusion, controlnets, everything. Ultimately the fucking sketch feature and drawing a giant pickle did it.

There is no sense of imminent_ in the models. You either are, or are not. If you try to prompt this you'll get weird nightmare penises stuck into nothing.

Anonymous

Anonymous Sun 21 May 2023 15:33:32 No.7387762 Report

Quoted By:

>>7387747
Thanks

Anonymous

Anonymous Sun 21 May 2023 15:36:10 No.7387767 Report

Quoted By: >>7387770

>>7387757
>look elf I'm pickle dick

Anonymous

Anonymous Sun 21 May 2023 15:37:46 No.7387770 Report

Quoted By:

>>7387767
i've almost died due to laugh with hot tea in my mouth
i fucking hate you

Anonymous

View Same Google ImgOps iqdb SauceNAO catbox_ojv2kt.png, 1MiB, 928x1280

Anonymous Sun 21 May 2023 15:53:11 No.7387778 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO 00297-2855657965.png, 2MiB, 1152x1664

Anonymous Sun 21 May 2023 15:53:33 No.7387781 Report

Quoted By: >>7387783 >>7387784 >>7387794 >>7387825 >>7387892 >>7387952

>https://mega.nz/folder/ZvA21I7L#ZZzU42rdAyWFOWQ_O94JaQ/folder/0io1WQTa
>>7387622
Fuck it, you guys can have it.

Here's a panties/leotard aside sex lora. It's good enough in my book, just gonna pretend it was only for sex positions all along.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00253-1116451866.png, 2MiB, 1152x1664

Anonymous Sun 21 May 2023 15:54:40 No.7387783 Report

Quoted By: >>7387952

>>7387781

Anonymous

Anonymous Sun 21 May 2023 15:54:55 No.7387784 Report

Quoted By:

>>7387781
based

Anonymous

View Same Google ImgOps iqdb SauceNAO 00134-3879998650-nsfw, scan, ((( (...).png, 2MiB, 1024x1536

Anonymous Sun 21 May 2023 16:02:04 No.7387794 Report

Quoted By: >>7387968

>>7387781
Based, thanks.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00360-2917561967.png, 2MiB, 1024x1536

Anonymous Sun 21 May 2023 16:06:19 No.7387801 Report

Quoted By: >>7388580

Anonymous

Anonymous Sun 21 May 2023 16:09:55 No.7387812 Report

Quoted By:

>>7387406
>"Did Sephiroth do this?"

Anonymous

Anonymous Sun 21 May 2023 16:15:34 No.7387820 Report

Quoted By: >>7387841

>>7386428
Care to share when you finish anon?

Anonymous

Anonymous Sun 21 May 2023 16:21:21 No.7387825 Report

Quoted By: >>7387896

>>7387781
there's no safetensors anon

Anonymous

Anonymous Sun 21 May 2023 16:26:38 No.7387832 Report

Quoted By: >>7387837

>>7387490
I am really tempted to try to make an ecchi dungeon crawler VN.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00406-1451876591.png, 1MiB, 1024x1536

Anonymous Sun 21 May 2023 16:27:52 No.7387835 Report

Quoted By:

Anonymous

Anonymous Sun 21 May 2023 16:29:50 No.7387837 Report

Quoted By:

>>7387832
you won't

you're like the guy from welcome to the nhk

Anonymous

View Same Google ImgOps iqdb SauceNAO catbox_4gdmhl.png, 1MiB, 928x1280

Anonymous Sun 21 May 2023 16:31:53 No.7387839 Report

Quoted By: >>7387841 >>7387876 >>7387965

Here's your msgk shortstack fix, Bebebe LoCon:

https://mega.nz/folder/yBg2WSiZ#WcJDDoyjhM6WnGobsKhiHQ

May redo it later but I think its good enough as of now, been using the last epoch mostly but lower ones work fine if you want to preserve the model style.

Anonymous

View Same Google ImgOps iqdb SauceNAO catbox_wsdp6f.png, 1MiB, 1024x1280

Anonymous Sun 21 May 2023 16:33:06 No.7387841 Report

Quoted By:

>>7387839

for

>>7387820

Anonymous

View Same Google ImgOps iqdb SauceNAO 00419-1465556594.png, 2MiB, 1024x1536

Anonymous Sun 21 May 2023 16:40:45 No.7387854 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO 01126-2739438440.png, 1MiB, 1048x1568

Anonymous Sun 21 May 2023 16:42:56 No.7387858 Report

Quoted By: >>7387924

Anonymous

Anonymous Sun 21 May 2023 16:49:32 No.7387876 Report

Quoted By: >>7387889

>>7387839
nice style but seems to be majorly overfit in regards to clothing and colour, favouring sleeveless clothes and bright colours.
I guess that's just what the guy draws.

Anonymous

Anonymous Sun 21 May 2023 16:53:40 No.7387882 Report

Quoted By:

>>7387701
FYI sdp and new xformers are deterministic.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00470-2777934934.png, 1MiB, 1024x1536

Anonymous Sun 21 May 2023 16:56:34 No.7387885 Report

Quoted By:

How the fuck do elf ears work

Anonymous

View Same Google ImgOps iqdb SauceNAO 00164-826930915.png, 1MiB, 1024x1280

Anonymous Sun 21 May 2023 16:57:46 No.7387889 Report

Quoted By:

>>7387876
does sleeves just fine though, and yes because its one of just a few characters he draws constantly

Anonymous

Anonymous Sun 21 May 2023 16:58:01 No.7387892 Report

Quoted By: >>7387896

>>7387781
I want the hiramoto anon to post his prototype manga lora

Anonymous

View Same Google ImgOps iqdb SauceNAO image.jpg, 2MiB, 2880x4160

Anonymous Sun 21 May 2023 16:59:11 No.7387896 Report

Quoted By: >>7387903 >>7387933

(You) should try out tiled noise inversion upscaling.

I did clean up the small details on this a little bit with lama cleaner because the original image was actually pretty dirty/noisy in the first place, but the majority of it is original.

>>7387825
Uploaded, previous upload failed for some reason.

>>7387892
Wrong anon? Not sure what the context of your reply is.

Anonymous

Anonymous Sun 21 May 2023 17:03:16 No.7387903 Report

Quoted By:

>>7387896
so he sees my reply and considers releasing it.

Anonymous

Anonymous Sun 21 May 2023 17:06:11 No.7387905 Report

Quoted By:

For whatever reason, at higher resolutions, it loves making roast beef vaginas.

Anonymous

View Same Google ImgOps iqdb SauceNAO 01301-2631447233.png, 1MiB, 1536x1024

Anonymous Sun 21 May 2023 17:08:50 No.7387908 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO 9996490989765858-1421578478-kill (...).png, 3MiB, 1600x2400

Anonymous Sun 21 May 2023 17:11:27 No.7387909 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO 01420-1918309800.png, 409KiB, 768x512

Anonymous Sun 21 May 2023 17:20:55 No.7387913 Report

Quoted By:

not final

Anonymous

View Same Google ImgOps iqdb SauceNAO 00068-4066032788.png, 2MiB, 1024x1536

Anonymous Sun 21 May 2023 17:27:42 No.7387921 Report

Quoted By:

Final.

Anonymous

Anonymous Sun 21 May 2023 17:27:56 No.7387922 Report

Quoted By: >>7387923

if only we had more loras like horsefucker-v1-512-epoch4...

Anonymous

Anonymous Sun 21 May 2023 17:30:26 No.7387923 Report

Quoted By:

>>7387922
we just can't deal with it without the furgods to guide us...

Anonymous

Anonymous Sun 21 May 2023 17:31:31 No.7387924 Report

Quoted By: >>7388062

>>7387858
That's quite nice, minus the triple belly button
Catbox?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00567-3429770243.png, 2MiB, 1024x1536

Anonymous Sun 21 May 2023 17:32:46 No.7387927 Report

Quoted By: >>7388396

Anonymous

View Same Google ImgOps iqdb SauceNAO 00004-3472842343.png, 1MiB, 1024x1024

Anonymous Sun 21 May 2023 17:36:16 No.7387931 Report

Quoted By: >>7387932

Liking the bebe lora, thanks anon. I still can't figure out hands though.

Anonymous

Anonymous Sun 21 May 2023 17:38:09 No.7387932 Report

Quoted By:

>>7387931
For my lax hand standards I would say that this is acceptable.

Anonymous

Anonymous Sun 21 May 2023 17:38:27 No.7387933 Report

Quoted By: >>7388010

>>7387896
>tiled noise inversion upscaling
Huh?

Anonymous

View Same Google ImgOps iqdb SauceNAO xyz_grid-0054-944740109-solo, Na (...).jpg, 324KiB, 4000x409

Anonymous Sun 21 May 2023 17:38:45 No.7387934 Report

Quoted By: >>7387938 >>7387942

...
yeah ok im just going to start removing files to ese which fucking image is causing this absolute retardation. this is insane

Anonymous

Anonymous Sun 21 May 2023 17:40:36 No.7387935 Report

Quoted By:

hahah

Anonymous

Anonymous Sun 21 May 2023 17:43:21 No.7387938 Report

Quoted By:

>>7387934
must be some image with tiles or some pattern

Anonymous

View Same Google ImgOps iqdb SauceNAO 00606-2223358281.png, 2MiB, 1024x1536

Anonymous Sun 21 May 2023 17:44:06 No.7387939 Report

Quoted By: >>7387941

Anonymous

Anonymous Sun 21 May 2023 17:47:05 No.7387941 Report

Quoted By: >>7387946

>>7387939
gato

Anonymous

Anonymous Sun 21 May 2023 17:47:09 No.7387942 Report

Quoted By: >>7387945

>>7387934
what's supposed to be the problem here?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00616-1915708466.png, 2MiB, 1024x1536

Anonymous Sun 21 May 2023 17:48:01 No.7387943 Report

Quoted By:

Anonymous

Anonymous Sun 21 May 2023 17:49:21 No.7387945 Report

Quoted By:

>>7387942
shes supposed to be sitting on stairs but it degenerates them into wooden tiles or chairs or whatever that is. The rightmost ones get it right, rest doesnt.
Thing is, the images are the same, its just using old tags from v1 tagger and no matter how else I tag this, whether with cleaned 0.15 or on 0.4 confidence it keeps happening. Pure insanity

Anonymous

Anonymous Sun 21 May 2023 17:49:27 No.7387946 Report

Quoted By: >>7387948 >>7388060

>>7387941
https://files.catbox.moe/j2cgsc.png

Anonymous

Anonymous Sun 21 May 2023 17:51:38 No.7387948 Report

Quoted By:

>>7387946
>that prompt
At least I'll check out the model!

Anonymous

Anonymous Sun 21 May 2023 17:53:44 No.7387951 Report

Quoted By: >>7387953

What dictates model switching speed? GPU or drive speed?

Anonymous

View Same Google ImgOps iqdb SauceNAO 02531-3751756953.png, 1MiB, 1024x1024

Anonymous Sun 21 May 2023 17:56:32 No.7387952 Report

Quoted By: >>7387991 >>7388018

>>7387781
>>7387783
i tried with no negative prompt and got great results. do we even need negative prompts anymore?
>raw output no upscaling

Anonymous

Anonymous Sun 21 May 2023 17:56:32 No.7387953 Report

Quoted By: >>7387954 >>7387977

>>7387951
drive/cpu , you can cache it on RAM though and switch between a few

Anonymous

Anonymous Sun 21 May 2023 17:56:59 No.7387954 Report

Quoted By: >>7387958

>>7387953
How do I do that?

Anonymous

Anonymous Sun 21 May 2023 17:58:25 No.7387957 Report

Quoted By: >>7387959

I'm so pissed, I'm very close to my image being done but every time I go to upscale it it gets seriously mangled.

Anonymous

Anonymous Sun 21 May 2023 17:58:26 No.7387958 Report

Quoted By: >>7387977

>>7387954
Settings > Stable Diffusion, top sliders

Anonymous

Anonymous Sun 21 May 2023 17:58:58 No.7387959 Report

Quoted By: >>7387974

>>7387957
surely you are not using latent

Anonymous

View Same Google ImgOps iqdb SauceNAO 00005-3179040215.png, 1007KiB, 768x1152

Anonymous Sun 21 May 2023 17:58:58 No.7387960 Report

Quoted By:

I was excited to boost it with the pailoli lora but it doesn't mix well, fries pretty bad at full weight. Kind of works diluted to get the base proportions and then i2i to rewrite bebe's style.

Anonymous

Anonymous Sun 21 May 2023 17:59:52 No.7387961 Report

Quoted By:

>>7387600
King shit

Anonymous

Anonymous Sun 21 May 2023 18:00:26 No.7387962 Report

Quoted By: >>7387967 >>7388429 >>7388443

please let batch 1 be shit so i don't feel compelled to rebake 30 loras again

Anonymous

Anonymous Sun 21 May 2023 18:02:17 No.7387965 Report

Quoted By: >>7387970

>>7387839
is this a lora or lycoris

Anonymous

Anonymous Sun 21 May 2023 18:03:01 No.7387967 Report

Quoted By: >>7387972

>>7387962
everytime I tried raising from batch1 I had worse results, good luck.

Anonymous

Anonymous Sun 21 May 2023 18:03:01 No.7387968 Report

Quoted By: >>7388012

>>7387794
catbox please

Anonymous

View Same Google ImgOps iqdb SauceNAO Another Round.jpg, 244KiB, 1024x1536

Anonymous Sun 21 May 2023 18:03:02 No.7387969 Report

Quoted By: >>7388452

Anonymous

Anonymous Sun 21 May 2023 18:03:05 No.7387970 Report

Quoted By: >>7388089

>>7387965
try reading

Anonymous

Anonymous Sun 21 May 2023 18:04:23 No.7387972 Report

Quoted By: >>7387975

>>7387967
I was also meaning to check out adafactor...

Anonymous

View Same Google ImgOps iqdb SauceNAO 81e17467bf0cc6bfdeef30823b3fb9d0.png, 495KiB, 414x493

Anonymous Sun 21 May 2023 18:04:59 No.7387973 Report

Quoted By: >>7388020 >>7389001

>update to adetailer
>pic related happens all the time now

Anonymous

View Same Google ImgOps iqdb SauceNAO 02108-3839156849.png, 1MiB, 1536x1024

Anonymous Sun 21 May 2023 18:05:16 No.7387974 Report

Quoted By: >>7387976

>>7387959
No, it's just that little details get fucked up so bad and I spend forever trying to repair them.

Anonymous

Anonymous Sun 21 May 2023 18:06:02 No.7387975 Report

Quoted By: >>7387980

>>7387972
prob not worth it, you're better off checking dadaptation with the proper commands added to json - afaik everyone who used easyscripts wasnt using it properly cause it doesnt add them automatically.

Anonymous

Anonymous Sun 21 May 2023 18:06:39 No.7387976 Report

Quoted By: >>7387981 >>7387993

>>7387974
Anon, what kind of fucking upscaler are you using, god damn?
This looks like ESRGAN, don't use that.

Anonymous

Anonymous Sun 21 May 2023 18:06:44 No.7387977 Report

Quoted By: >>7387982

>>7387953
>>7387958
That actually slows down my switching speed drastically, from 2 seconds to 30 seconds

Anonymous

Anonymous Sun 21 May 2023 18:06:48 No.7387978 Report

Quoted By:

So I ended up biting the bullet and grabbed controlnet but not sure what I'm doing wrong since it's not even trying the poses and instead creates few black images after the actual one.

No errors either

Anonymous

Anonymous Sun 21 May 2023 18:07:41 No.7387980 Report

Quoted By: >>7387990

>>7387975
Proper commands as in? I'm the OG dadapt anon so I think it has been working so far.

Anonymous

Anonymous Sun 21 May 2023 18:08:31 No.7387981 Report

Quoted By: >>7387984

>>7387976
What should be used?

Anonymous

Anonymous Sun 21 May 2023 18:09:17 No.7387982 Report

Quoted By: >>7387983

>>7387977
Turns out you are not allowed to set VAE to auto when caching checkpoints, otherwise it will reload them every time you switch

Anonymous

Anonymous Sun 21 May 2023 18:10:43 No.7387983 Report

Quoted By: >>7387986

>>7387982
you shouldnt be using auto VAE ever to begin with

Anonymous

Anonymous Sun 21 May 2023 18:10:43 No.7387984 Report

Quoted By: >>7387985 >>7387992 >>7387993

>>7387981
I say none if you're upscaling via img2img and have the VRAM, but if you do have to use the extensions, use animesharp. Or remacri? It's okay.

Anonymous

Anonymous Sun 21 May 2023 18:11:43 No.7387985 Report

Quoted By: >>7387995

>>7387984
>upscaling via img2img
pretty sure that uses an upscaler? I've been told so

Anonymous

Anonymous Sun 21 May 2023 18:11:51 No.7387986 Report

Quoted By:

>>7387983
I know, I was reinstalling the WebUI, that's why it was still set.

Does it not work with the amount to cache set to 1? It only seems to work with atleast 2

Anonymous

View Same Google ImgOps iqdb SauceNAO toSlap-or-not-toSlap.webm, 1MiB, 512x683

Anonymous Sun 21 May 2023 18:12:33 No.7387989 Report

Quoted By:

Anonymous

Anonymous Sun 21 May 2023 18:12:43 No.7387990 Report

Quoted By: >>7387995

>>7387980
https://rentry.org/59xed3#dadaptation the arguments listed here

Anonymous

Anonymous Sun 21 May 2023 18:14:29 No.7387991 Report

Quoted By:

>>7387952
Very nice.

Anonymous

Anonymous Sun 21 May 2023 18:14:52 No.7387992 Report

Quoted By: >>7387995

>>7387984
img2img is what's fucking my shit up to begin with

Anonymous

Anonymous Sun 21 May 2023 18:15:17 No.7387993 Report

Quoted By: >>7387995

>>7387976
>>7387984
Animesharp and Remacri are ESRGAN upscalers, what do you even mean?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00032-388513-1girl, hair ornamen (...).png, 1MiB, 1280x1792

Anonymous Sun 21 May 2023 18:16:24 No.7387995 Report

Quoted By: >>7387999 >>7388000

Huh, looks like I'm gonna be using Yabal3 now. Really nice style.
>>7387985
That's what I use. Just set the resolution to what you want and the proper denoise. It's weird because highresfix without an upscaler sucks, but this is good. You can also do latent via img2img if you want.
It uses appropriately more VRAM, but you don't get tiling problems.
>>7387990
I set up my script a while ago and it did that, so yes. I just used cosine with retards instead.
>>7387992
Well yeah, you were using a shit upscaler, if that was you.
>>7387993
You didn't know that you can use upscalers in highresfix and sd upscale?

Anonymous

View Same Google ImgOps iqdb SauceNAO 02136-2132094977.jpg, 242KiB, 3072x2048

Anonymous Sun 21 May 2023 18:17:23 No.7387997 Report

Quoted By:

Anonymous

Anonymous Sun 21 May 2023 18:18:12 No.7387999 Report

Quoted By: >>7388001

>>7387995
>You didn't know that you can use upscalers in highresfix and sd upscale?
?????????????
Huh? Yeah I knew but Remacri and Animesharp are ESRGAN upscalers

Anonymous

Anonymous Sun 21 May 2023 18:18:41 No.7388000 Report

Quoted By: >>7388001

>>7387995
Settings>Upscaling>Upscaler for img2img

Anonymous

Anonymous Sun 21 May 2023 18:19:19 No.7388001 Report

Quoted By: >>7388004

>>7387999
Yes?
>>7388000
???

Anonymous

Anonymous Sun 21 May 2023 18:19:31 No.7388003 Report

Quoted By: >>7388009

>have a seed locked in
>large breasts in prompt
>hands are perfect
>change to medium breast prompt
>hands turn into monsters

Anonymous

Anonymous Sun 21 May 2023 18:19:54 No.7388004 Report

Quoted By: >>7388005

>>7388001
That's where the upscaler is set when you increase image size on img2img

Anonymous

Anonymous Sun 21 May 2023 18:20:23 No.7388005 Report

Quoted By: >>7388008

>>7388004
Oh wait, I get it now. LMAO! Never knew this was what I was using. Picrel was animesharp then.

Anonymous

View Same Google ImgOps iqdb SauceNAO firefox_HK3cCxnUMT.png, 10KiB, 857x75

Anonymous Sun 21 May 2023 18:21:52 No.7388008 Report

Quoted By: >>7388013 >>7388340

>>7388005
This should be mandatory for everyone using a1111

Anonymous

Anonymous Sun 21 May 2023 18:21:54 No.7388009 Report

Quoted By:

>>7388003
Her hands are coming to strangle you for trying to take away her pride and joy.

Anonymous

View Same Google ImgOps iqdb SauceNAO 1.png, 66KiB, 835x723

Anonymous Sun 21 May 2023 18:22:56 No.7388010 Report

Quoted By: >>7389193

>>7387933
It's just combining multidiffusion's noise inversion with controlnet's tile model. Solves the issue of "dirty" details. miniature people, etc when upscaling at >0.4 denoise.

You'll still get a high amount of details without having to deal with any of the weird shit. Not sure if this is achievable with ultimate sd upscaling, as I haven't tried it.

Anonymous

Anonymous Sun 21 May 2023 18:23:52 No.7388012 Report

Quoted By: >>7388037

>>7387968
https://files.catbox.moe/2bmza5.png

Anonymous

Anonymous Sun 21 May 2023 18:25:04 No.7388013 Report

Quoted By: >>7388017

>>7388008
I must've forgotten about it since it was set to animesharp in the first place. Funny tho.
And do you actually use the fix steps? I haven't found that much use for it myself.

Anonymous

Anonymous Sun 21 May 2023 18:25:57 No.7388017 Report

Quoted By: >>7388019

>>7388013
Only for upscaling, as this makes the steps behave as they do for hires fix

Anonymous

Anonymous Sun 21 May 2023 18:26:01 No.7388018 Report

Quoted By:

>>7387952
Depends, negatives push it towards a particular artstyle and "clean" look but results are decent enough without negatives.

Anonymous

Anonymous Sun 21 May 2023 18:26:37 No.7388019 Report

Quoted By: >>7388021

>>7388017
I know, but I'm asking whether you found it to be desirable in the first place.
I haven't done extensive enough testing to know.

Anonymous

Anonymous Sun 21 May 2023 18:28:10 No.7388020 Report

Quoted By: >>7388024

>>7387973
Mine is still working. Are you using it together with other extensions? Regional prompting absolutely fucks up adetailer.

Anonymous

Anonymous Sun 21 May 2023 18:28:28 No.7388021 Report

Quoted By: >>7388022 >>7388023

>>7388019
I've never really needed more than 10 steps for upscaling, but I sometimes increase or decrease the denoise, so it's easier to fix the steps than to do 20*denoisingStrength every time

Anonymous

Anonymous Sun 21 May 2023 18:29:28 No.7388022 Report

Quoted By:

>>7388021
10/denoisingStrength actually, I'm retarded

Anonymous

Anonymous Sun 21 May 2023 18:30:42 No.7388023 Report

Quoted By: >>7388031

>>7388021
There is a bit more detail if you use it over using the lowered step amounts. Why not.
And I wish they finally separated img2img and inpainting tabs for stuff like that and scripts.

Anonymous

Anonymous Sun 21 May 2023 18:30:52 No.7388024 Report

Quoted By: >>7388032

>>7388020
What the fuck, why can't it handle openpose? Turned it off and now it works.
It worked before the update..

Anonymous

Anonymous Sun 21 May 2023 18:34:27 No.7388031 Report

Quoted By:

>>7388023
(Though the difference at least on on 0.35 and 0.5 between the 6 adjusted steps and 15 fixed ones is near imperceptible if you werent comparing side by sides.)

Anonymous

Anonymous Sun 21 May 2023 18:34:32 No.7388032 Report

Quoted By:

>>7388024
If you're on a newer version of WebUI you have backup hashes of your extensions prior to the pull stored automatically. Try reverting from the backup/restore tab until they fix it.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00091-3701222264.png, 1MiB, 1024x1280

Anonymous Sun 21 May 2023 18:38:19 No.7388036 Report

Quoted By: >>7388040 >>7388349

So what are the current cloudshit options currently? What are poorbros using? I don't mind paying reasonable amount and I don't think I'm gonna be prompting too much.
Lora training still can be done on colab and I don't make many loras anyway, so I just want to prompt and spare my GPU.

Anonymous

Anonymous Sun 21 May 2023 18:38:55 No.7388037 Report

Quoted By:

>>7388012
Bless

Anonymous

View Same Google ImgOps iqdb SauceNAO catbox_de82rx.png, 3MiB, 1664x1664

Anonymous Sun 21 May 2023 18:39:22 No.7388039 Report

Quoted By:

>>7385364
First time trying to train a character with multiple outfits, this advice was pretty good. I can easily prompt the different outfits with positive or negative prompts with a bit a gacha. Might retrain some other characters I did in the past. Thanks anon.

Anonymous

Anonymous Sun 21 May 2023 18:40:05 No.7388040 Report

Quoted By: >>7388048

>>7388036
gatoboxxu
Also, the wiki has some options. I think most are using paperspace.
Runpod is cheapo, but it's a pay-as-you-go and I've had shit experiences with it before.

Anonymous

View Same Google ImgOps iqdb SauceNAO 02137-2928663215.png, 2MiB, 1024x2048

Anonymous Sun 21 May 2023 18:40:42 No.7388043 Report

Quoted By:

I can't get Tile Diffusion to work as it should but I can use it for instant loss.

Anonymous

Anonymous Sun 21 May 2023 18:41:05 No.7388044 Report

Quoted By:

How do I stop controlnet from chopping heads off? I've included everything from face to eyes in the prompt but no luck

Anonymous

Anonymous Sun 21 May 2023 18:42:33 No.7388046 Report

Quoted By: >>7388049

so run into a bit of a problem trying to train at 1024 size images, im not sure if its because im using min_snr_gamma but even reducing the learning rate by a factor of 10 causes it to overbake like crazy to the point that when genin any weight about .1 just produces noise

Anonymous

View Same Google ImgOps iqdb SauceNAO 00001-2870825446.png, 2MiB, 1024x1536

Anonymous Sun 21 May 2023 18:43:04 No.7388047 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO 00058-2635972938.png, 1MiB, 1024x1280

Anonymous Sun 21 May 2023 18:44:02 No.7388048 Report

Quoted By: >>7388299

>>7388040
https://files.catbox.moe/zs9po1.png

>Runpod
>I've had shit experiences with it before.
Yeah that's what it looks like. I don't want to get thrown out from the session. Ideally it would be a subscription service where you pay for a month and there's reasonable GPU availability. Guess I'll try paperspace

Anonymous

Anonymous Sun 21 May 2023 18:44:09 No.7388049 Report

Quoted By: >>7388055 >>7388056 >>7388335

>>7388046
I've baked at 1024 before but I don't use that gamma, so that's probably it?
Can't really recommend it.

Anonymous

Anonymous Sun 21 May 2023 18:46:44 No.7388055 Report

Quoted By: >>7388335

>>7388049
>I've baked at 1024 before but I don't use that gamma, so that's probably it?
seems not cause problems at other sizes,
at 768 i didn't see any funny business

also example xy plot
https://files.catbox.moe/1shzkb.png

Anonymous

Anonymous Sun 21 May 2023 18:47:25 No.7388056 Report

Quoted By: >>7388113

>>7388049
>Can't really recommend it.
people were saying it allows models to converge up to three times faster

Anonymous

View Same Google ImgOps iqdb SauceNAO 00756-1143599829.png, 1MiB, 1536x1024

Anonymous Sun 21 May 2023 18:47:31 No.7388057 Report

Quoted By:

Anonymous

Anonymous Sun 21 May 2023 18:52:03 No.7388059 Report

Quoted By: >>7388066 >>7388068

Does training at 1024 actually improve things, or are people just training them to flex on us vramlets?

Anonymous

Anonymous Sun 21 May 2023 18:52:17 No.7388060 Report

Quoted By:

>>7387946
Where do I get the flare lora? It's not listed in the gayshit.

Anonymous

Anonymous Sun 21 May 2023 18:52:23 No.7388061 Report

Quoted By:

correct me if im wrong but it seems to me that you have a lot more room before a lora overbakes if you are just training the unet right?

Anonymous

Anonymous Sun 21 May 2023 18:52:38 No.7388062 Report

Quoted By:

>>7387924
https://files.catbox.moe/3yd6w9.png
before moderate inpainting

Anonymous

View Same Google ImgOps iqdb SauceNAO 00188-3921403322-), dark skin, d (...).png, 1MiB, 1280x1408

Anonymous Sun 21 May 2023 18:53:49 No.7388063 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO xyz_grid-0061-212144573-solo, Na (...).jpg, 364KiB, 4000x493

Anonymous Sun 21 May 2023 18:54:19 No.7388064 Report

Quoted By:

...
I think it's improved alot since I removed https://files.catbox.moe/ofumfd.jpg
bizarre isnt it? But then again, the autotagger claimed this had "shoes off" or whatever so im not really sure what AI was seeing here.
Time to do a proper rebake sometime later.

Anonymous

Anonymous Sun 21 May 2023 18:56:10 No.7388066 Report

Quoted By: >>7388068

>>7388059
for me im doing this because i have a particular dataset due to the anime being something from the early 00s which lead to
lots of low quality low resolution images from the anime, and a handful of high quality high resolution images from the promotion material, artbooks ect..
my attempt is to train the lower resolution images , then resume training on the high quality images and my hope that at higher resolution more of the images layers are train to correct what was learned from the low resolution images.

Anonymous

Anonymous Sun 21 May 2023 19:00:22 No.7388068 Report

Quoted By: >>7388069 >>7388073

>>7388066
what is this thought process? how dumb do you have to be to come up with this lol
Making fun of you aside, just upscale the images at 1.1 through the extras tab - that should remove most graphical artifacts.

>>7388059
Dunno. Doubtful? Last time someone claimed its better and posted an x/y it really wasnt.
Truth be told [spoiler:lit]the difference between 512 and 768 isnt neccessarily there because of the resolutions - I certainly cant tell if a well trained lora is 512, but if someone asks I'll defend 768 placebo and die on this hill.[/spoiler:lit]

Anonymous

Anonymous Sun 21 May 2023 19:01:23 No.7388069 Report

Quoted By:

>>7388068
>realize /h/ doesnt allow funny black bar just as I click post
its so fucking over doomerbros you were right all along... /trash/ won.....

Anonymous

View Same Google ImgOps iqdb SauceNAO 00002-1930607733.0.png, 1MiB, 1024x1536

Anonymous Sun 21 May 2023 19:04:41 No.7388071 Report

Quoted By: >>7388074 >>7388216 >>7389117

>decides to train a lora to gen waifu
>using 25 SFW artworks of waifu as dataset and ignoring "at least 50-150 images" from guide because lazyness
>leaving all tags intact (yes, "highres" and "commentary request" too) BUT deleting artist tags because autism
>applying same tags to several different images because autism
>put them in 10_ and 20_ repeats folders with no quality pattern
>copying arguments from guide and not reading what are they doing
>training this abomination up to e10
>loading up e10 and slaps "hiiragi kagami, ryouou school uniform" in prompt box
>not only getting waifu perfectly, but also getting cool ass sketch style if applying style tags
>proompting for next 6 hours and having best nuts in my life

>in post-nut clarity decides to retrain it
>put myself through diaper, scat, vore, guro nightmare while sifting through 18k+ scrapped images to gather proper dataset (427 instead of 25)
>pruning all unneeded tags
>put dataset in 1_ to 5_ repeat folders while sticking to quality pattern
>trains it up to e10
>excitedly loading up e10 and slaps same prompt
>undercooked as fuck because of much lower repeats

>decides to retrain it again after asking /hdg/ about best training practicies
>dataset consists of 72 curated images that are extremely similar to first dataset
>erasing artist' signatures from them and some background elements to reduce amount of noise
>applying and pruning tags in a way to be able to easily get waifu default look and outfit by only prompting her name
>put them in 10_ and 20_ repeats folders while sticking to quality pattern
>trains it up to e10
>loading them into matrix and slaps same prompt
>e1-4 are undercooked, e5-10 are overcooked
>experiment with e4 on lower weights
>cool ass sketch style is nowhere to be seen, waifu looks like blurry shit

>puts "embarrassment" in prompt while testing all loras including first one
>first lora's picrel perfectly encapsulates AI feelings about this greentext

im done

inb4 "i aint reading that"

Anonymous

Anonymous Sun 21 May 2023 19:06:48 No.7388073 Report

Quoted By: >>7388077

>>7388068
>what is this thought process?
i don't think you are quit aware of the shortcuts and corner cutting early anime used.

Anonymous

Anonymous Sun 21 May 2023 19:07:30 No.7388074 Report

Quoted By: >>7388255

>>7388071
way too fucking long buddy, what the fuck.
you seem dumb so the results are about what one would expect
imagine even ending up with booru tags somehow, imagine using scrappers lmao

Anonymous

Anonymous Sun 21 May 2023 19:09:23 No.7388077 Report

Quoted By:

>>7388073
wait, by low quality you mean low quality frames where characters are off model? why would you use these? Any old anime had plenty good frames that weren't distorted to shit. Unsure what else could you mean by corner cutting if this aint it.

Anonymous

Anonymous Sun 21 May 2023 19:12:28 No.7388081 Report

Quoted By: >>7388087 >>7388117 >>7388123 >>7388160 >>7388232

Why is armpit a term everyone learns but humanity has never bothered to teach people the name for back of the knee or back of the elbow.
I can't prompt words no one has ever heard

Anonymous

Anonymous Sun 21 May 2023 19:17:51 No.7388087 Report

Quoted By:

>>7388081
its simple, everyone's secretly into armpits.

Anonymous

Anonymous Sun 21 May 2023 19:19:00 No.7388089 Report

Quoted By: >>7388100

>>7387970
I don't see either word in the post

Anonymous

Anonymous Sun 21 May 2023 19:22:25 No.7388100 Report

Quoted By: >>7388102

>>7388089
NTA but locon = lycoris

Anonymous

Anonymous Sun 21 May 2023 19:22:47 No.7388102 Report

Quoted By:

>>7388100
thank you

Anonymous

Anonymous Sun 21 May 2023 19:26:18 No.7388111 Report

Quoted By: >>7388116 >>7388132

>still using locons
anything the lycoris faggot made was scam

Anonymous

Anonymous Sun 21 May 2023 19:26:35 No.7388113 Report

Quoted By: >>7388118

>>7388056
Can't say it does? If anything wouldn't it be the opposite since people said the opposite about 768 compared to 1024?
Whatever's the case, I retrained over 10 at 1024 and ended up disappointed.

Anonymous

Anonymous Sun 21 May 2023 19:27:44 No.7388116 Report

Quoted By: >>7388126

>>7388111
Obsessed

Anonymous

Anonymous Sun 21 May 2023 19:27:48 No.7388117 Report

Quoted By:

>>7388081
kneepit ofc

Anonymous

Anonymous Sun 21 May 2023 19:28:49 No.7388118 Report

Quoted By: >>7388128

>>7388113
Compared to 512 I mean. People were saying how you need to train 768s for longer than 512. Funny how it's somehow the opposite now.

Anonymous

Anonymous Sun 21 May 2023 19:30:04 No.7388123 Report

Quoted By:

>>7388081
>humanity has never bothered to teach people the name for back of the knee or back of the elbow

Pretty sure it's called crook of the elbow/knee, anon.

Anonymous

Anonymous Sun 21 May 2023 19:30:46 No.7388126 Report

Quoted By: >>7388130

>>7388116
yes, im the obsessed guy for calling out the faggot that had to name a training method based on a fucking seasonal yuri show. Certainly he's not the obsessed one and not just a scammer clout chasing.

Anonymous

Anonymous Sun 21 May 2023 19:31:28 No.7388127 Report

Quoted By:

guys train yuruyuris now
the age of loras is over

Anonymous

Anonymous Sun 21 May 2023 19:31:47 No.7388128 Report

Quoted By: >>7388131

>>7388118
>Funny how it's somehow the opposite now.
its not and no one said it is.

Anonymous

Anonymous Sun 21 May 2023 19:32:17 No.7388130 Report

Quoted By: >>7388137

>>7388126
Oh it's actually named after that? Lol no wonder the acronym is so retarded

Anonymous

Anonymous Sun 21 May 2023 19:32:37 No.7388131 Report

Quoted By: >>7388136

>>7388128
>people were saying it allows models to converge up to three times faster
Haven't heard that myself tho.

Anonymous

Anonymous Sun 21 May 2023 19:32:41 No.7388132 Report

Quoted By: >>7388142

>>7388111
locons were a different branch, retard. Lycoris bullshit was later made and included loha/locon with it

Anonymous

Anonymous Sun 21 May 2023 19:32:54 No.7388134 Report

Quoted By: >>7388150

I have a very hard time remembering what naming is which for anything other than the original 'lora'.

Who the fuck even comes up with these names, man.

Anonymous

Anonymous Sun 21 May 2023 19:33:31 No.7388136 Report

Quoted By: >>7388141

>>7388131
He's talking about the gamma offset, dummy.

Anonymous

Anonymous Sun 21 May 2023 19:33:40 No.7388137 Report

Quoted By:

>>7388130
>Lol no wonder the acronym is so retarded
My favorite is the acronym for LION.

Anonymous

Anonymous Sun 21 May 2023 19:34:41 No.7388141 Report

Quoted By:

>>7388136
Oh then we both misunderstood each other because I was saying that I don't consider 1024 to be worth it, not gamma.

Anonymous

Anonymous Sun 21 May 2023 19:35:07 No.7388142 Report

Quoted By: >>7388146

>>7388132
it started with lycoris, locon was made much later on the same branch. why would you try to gaslight people in an oldfag thread?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00087-1423010696.png, 1MiB, 1024x1280

Anonymous Sun 21 May 2023 19:37:43 No.7388144 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO file.png, 385KiB, 1049x694

Anonymous Sun 21 May 2023 19:38:42 No.7388146 Report

Quoted By: >>7388158

>>7388142
https://github.com/KohakuBlueleaf/LyCORIS

Anonymous

Anonymous Sun 21 May 2023 19:41:54 No.7388150 Report

Quoted By: >>7388155 >>7388159

>>7388134
The person who came up with locon/lycoris literally wanted people to be training "LoLiCons", guess his plan fell thru somewhere
https://twitter.com/KBlueleaf/status/1630755151022813185

Anonymous

Anonymous Sun 21 May 2023 19:44:51 No.7388155 Report

Quoted By:

>>7388150
Good fucking thing his suggestion wasn't adopted. Get that shit outta here.

Anonymous

Anonymous Sun 21 May 2023 19:45:00 No.7388156 Report

Quoted By:

Did that one anon actually make a stubble lora?

Anonymous

Anonymous Sun 21 May 2023 19:45:17 No.7388158 Report

Quoted By: >>7388164 >>7388167 >>7388171

>>7388146
I dont care what that schizoid claims, Lycoris was the first thing that came out of his git. Locons came out later and had some more attention, newfag. Go look through the archive at the order of things

Anonymous

Anonymous Sun 21 May 2023 19:46:35 No.7388159 Report

Quoted By:

>>7388150
Kek, based

Anonymous

Anonymous Sun 21 May 2023 19:46:44 No.7388160 Report

Quoted By: >>7388166

>>7388081
https://danbooru.donmai.us/wiki_pages/kneepits?z=2

Are you retarded?

Anonymous

Anonymous Sun 21 May 2023 19:47:04 No.7388162 Report

Quoted By:

>>7387605
>>7387619
nice vags bro, how did you get them like that?

Anonymous

Anonymous Sun 21 May 2023 19:47:54 No.7388164 Report

Quoted By: >>7388178

>>7388158
>Locons came out later and had some more attention
backwards, locons were barely known, which is why you thing lycoris actually came first. Locons were around since way before, and the git repo/extension for them is also older than the lycoris one and actually integrates with the default lora tab unlike lycoris that adds a new one

Anonymous

Anonymous Sun 21 May 2023 19:48:32 No.7388166 Report

Quoted By:

>>7388160
Cool, I'll try that

Anonymous

Anonymous Sun 21 May 2023 19:49:00 No.7388167 Report

Quoted By:

>>7388158
check his twitter, the locon term was actually around before his obsession with the lycoris recoil anime.

Anonymous

Anonymous Sun 21 May 2023 19:49:37 No.7388171 Report

Quoted By:

>>7388158
>I dont care what that schizoid claims
The ironing

Anonymous

View Same Google ImgOps iqdb SauceNAO 00005-4028271458-1girl, mature f (...).png, 2MiB, 1280x1792

Anonymous Sun 21 May 2023 19:51:13 No.7388173 Report

Quoted By: >>7388199 >>7388298

doin your mom
doin doin your mom

Anonymous

Anonymous Sun 21 May 2023 19:52:41 No.7388178 Report

Quoted By:

>>7388164
>feb 28 for locon
>march 1 for lycoris
I'd have to go through the archives to make sure he didnt just remake his git but lazy, barely any difference anyways.
weird how the less meme shit had less attention if so, even if both suck

Anonymous

Anonymous Sun 21 May 2023 19:58:36 No.7388192 Report

Quoted By: >>7388196

does having xformers on make seeds not 100% the same? I guess I hadn't noticed til now, but I cannot get a seed to produce the same image anymore

Anonymous

View Same Google ImgOps iqdb SauceNAO 00274-3503787858.png, 2MiB, 1280x1408

Anonymous Sun 21 May 2023 20:00:26 No.7388196 Report

Quoted By: >>7388198 >>7388235

>>7388192
If using the xformers that comes with the webui, yes (0.0.17).

Since 0.0.19 and onwards, it is determistic.

Anonymous

Anonymous Sun 21 May 2023 20:01:59 No.7388198 Report

Quoted By: >>7388204

>>7388196
how the frock do I update it kind anon?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00017-990540820-1girl, mature fe (...).png, 2MiB, 1400x1960

Anonymous Sun 21 May 2023 20:02:30 No.7388199 Report

Quoted By:

>>7388173
Still doin her
Also fuck I hate what the AI does to sternums.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00870-2312539477.png, 1MiB, 1024x1536

Anonymous Sun 21 May 2023 20:05:45 No.7388203 Report

Quoted By: >>7388939

Anonymous

View Same Google ImgOps iqdb SauceNAO 00283-3463696343.png, 1MiB, 1280x1408

Anonymous Sun 21 May 2023 20:05:52 No.7388204 Report

Quoted By: >>7388214 >>7388222 >>7388231

>>7388198
You just enable your venv of your stable diffusion, and use "pip install xformers==0.0.19"

https://pypi.org/project/xformers/0.0.19/

Anonymous

Anonymous Sun 21 May 2023 20:11:57 No.7388214 Report

Quoted By: >>7388242

>>7388204
gato

Anonymous

Anonymous Sun 21 May 2023 20:12:40 No.7388216 Report

Quoted By:

>>7388071
autism

Anonymous

View Same Google ImgOps iqdb SauceNAO 00054-2589083158.png, 1MiB, 1024x1024

Anonymous Sun 21 May 2023 20:13:54 No.7388218 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO 01231-1698051495.png, 1MiB, 1024x1536

Anonymous Sun 21 May 2023 20:15:14 No.7388220 Report

Quoted By: >>7388225

Anonymous

Anonymous Sun 21 May 2023 20:16:09 No.7388222 Report

Quoted By: >>7388242

>>7388204
what do I open in the venv folder to input the pip install line at?

Anonymous

Anonymous Sun 21 May 2023 20:16:57 No.7388225 Report

Quoted By: >>7388229

>>7388220
gato too

Anonymous

Anonymous Sun 21 May 2023 20:18:05 No.7388229 Report

Quoted By:

>>7388225
https://files.catbox.moe/z0dn40.png

Anonymous

Anonymous Sun 21 May 2023 20:18:17 No.7388231 Report

Quoted By: >>7388242

>>7388204
I know what model this is. It shines even through character loras, fucking kek
>new xformers
does it even help performance or is this gonna cause vram issues like pulling

Anonymous

Anonymous Sun 21 May 2023 20:18:47 No.7388232 Report

Quoted By: >>7388238 >>7388240 >>7388244

>>7388081
kneepussy

Anonymous

Anonymous Sun 21 May 2023 20:19:33 No.7388234 Report

Quoted By:

>1560x2184 img2img possible
MORE

Anonymous

Anonymous Sun 21 May 2023 20:19:36 No.7388235 Report

Quoted By: >>7388242

>>7388196
catbox pl0x

Anonymous

Anonymous Sun 21 May 2023 20:19:59 No.7388238 Report

Quoted By:

>>7388232
knussy

Anonymous

Anonymous Sun 21 May 2023 20:20:49 No.7388240 Report

Quoted By:

>>7388232
based

Anonymous

View Same Google ImgOps iqdb SauceNAO 00292-2419605695.png, 2MiB, 1280x1408

Anonymous Sun 21 May 2023 20:21:11 No.7388242 Report

Quoted By: >>7388249 >>7388256 >>7388283

>>7388214
https://files.catbox.moe/7k51bs.png

>>7388222
On the sd webui main folder
.\venv\Scripts\activate
pip install xformers==0.0.19

>>7388231
Not sure, but you have to be prepared to maybe suffer errors is not using a recommended lib version that comes with the default webui. It can go without any issue either.

>>7388235
https://files.catbox.moe/ap7iy9.png

Anonymous

View Same Google ImgOps iqdb SauceNAO file.png, 12KiB, 600x639

Anonymous Sun 21 May 2023 20:21:26 No.7388244 Report

Quoted By:

>>7388232

Anonymous

Anonymous Sun 21 May 2023 20:22:43 No.7388249 Report

Quoted By: >>7388252

>>7388242
Oh it's that one model.
RELEASE IT i beg you

Anonymous

Anonymous Sun 21 May 2023 20:25:42 No.7388252 Report

Quoted By:

>>7388249
just mix the lycoris model with anything else, all I see in it are lycoreco girls though no matter how much he mixes it, so bad even chara loras dont salvage

Anonymous

Anonymous Sun 21 May 2023 20:26:42 No.7388255 Report

Quoted By: >>7388259 >>7388260

>>7388074
>way too long
go back you fucking zoomer

Anonymous

Anonymous Sun 21 May 2023 20:26:42 No.7388256 Report

Quoted By:

>>7388242
Ty anon

Anonymous

Anonymous Sun 21 May 2023 20:28:43 No.7388259 Report

Quoted By: >>7388262 >>7388278

>>7388255
maybe stop wasting people's time with your life story and get to the fucking point

Anonymous

Anonymous Sun 21 May 2023 20:28:53 No.7388260 Report

Quoted By: >>7388278

>>7388255
i've used up all available text space to post that shit so he got a point

Anonymous

Anonymous Sun 21 May 2023 20:29:14 No.7388262 Report

Quoted By: >>7388267

>>7388259
point: skill issue
p much

Anonymous

View Same Google ImgOps iqdb SauceNAO 00915-2621963860.png, 1MiB, 1024x1536

Anonymous Sun 21 May 2023 20:29:17 No.7388263 Report

Quoted By: >>7388939

Anonymous

Anonymous Sun 21 May 2023 20:30:23 No.7388265 Report

Quoted By: >>7388288

>>7387503
https://files.catbox.moe/h7kqqi.png

Anonymous

Anonymous Sun 21 May 2023 20:32:19 No.7388267 Report

Quoted By: >>7388268

>>7388262
yeah its a skill issue because you cannot write you autist

Anonymous

Anonymous Sun 21 May 2023 20:36:13 No.7388268 Report

Quoted By: >>7388484

>>7388267
anon...

Anonymous

View Same Google ImgOps iqdb SauceNAO 9996490989765870-2967771166-pura (...).png, 2MiB, 1024x1536

Anonymous Sun 21 May 2023 20:42:30 No.7388275 Report

Quoted By: >>7388551

Anonymous

Anonymous Sun 21 May 2023 20:43:39 No.7388278 Report

Quoted By: >>7388289

>>7388259
if you get bothered by a small amount of text your time is already a worthless waste, there's a different website with a smaller character limit more suited to your impressive 50 wpm zoomer skills

>>7388260
his only point is that he didn't learn how to read until 9th grade, assuming he got that far

Anonymous

Anonymous Sun 21 May 2023 20:45:13 No.7388280 Report

Quoted By:

>>7387499
the real protip is using chatGPT to LEARN coding alongside a bonafide guide, and then once you know language+framework of choice, using chatGPT to increase your productivity. there's so many fucking niche questions i've asked chatGPT and gotten instant responses on that could be verified once I knew the right language that took HOURRRRS of searching to finally find the right vocab to ask the question properly to then find and understand the answer. total game changer

Anonymous

Anonymous Sun 21 May 2023 20:47:14 No.7388283 Report

Quoted By:

>>7388242
my webui build must be held together with tape at this point because i keep breaking it and fixing it to the point where i get a ton of errors on launch but it works anyway
off to update my torch and python so i can use this

Anonymous

Anonymous Sun 21 May 2023 20:48:27 No.7388288 Report

Quoted By:

>>7388265
bless your soul

Anonymous

Anonymous Sun 21 May 2023 20:48:37 No.7388289 Report

Quoted By: >>7388305 >>7388481

>>7388278
you must be fucking retarded if you're willing to read pointlessly long life stories like this. He could have made that WAY shorter without losing any piece of information. Its not funny, its not interesting, its the equivalent of me telling you everything I've done tonight before actually answering your post.
Its not about being a zoomer or having attention span issues, its just fucking lines of white noise, its offensive to anyone actually using his brain.
And no, my time is not worthless, what kinda fucking argument is this even, "you don't want to waste time on pointless shit so your time is worthless"? You're so fucking retarded the mere fact you mention schooling is offensive. Christ.

Anonymous

Anonymous Sun 21 May 2023 20:49:49 No.7388292 Report

Quoted By:

>>7387499
i never tried chatGPT for code but i did try github co pilot and that ended up biting me in the ass.

Anonymous

Anonymous Sun 21 May 2023 20:52:11 No.7388296 Report

Quoted By:

Has anyone used this?
https://github.com/s1dlx/sd-webui-bayesian-merger/wiki/Run

I was pointed to it but I can't make heads or tails of it

Anonymous

View Same Google ImgOps iqdb SauceNAO yoomoo.png, 2MiB, 1280x1536

Anonymous Sun 21 May 2023 20:52:43 No.7388298 Report

Quoted By: >>7388551

>>7388173
what's happening forum

Anonymous

Anonymous Sun 21 May 2023 20:55:00 No.7388299 Report

Quoted By:

>>7388048
thanks for catbox kind prompter

Anonymous

Anonymous Sun 21 May 2023 21:00:02 No.7388305 Report

Quoted By:

>>7388289
Disgusting cope
>lines of white noise
just like your entire post

Anonymous

Anonymous Sun 21 May 2023 21:02:39 No.7388306 Report

Quoted By:

go back to sleep nafags and leave this thread be

Anonymous

View Same Google ImgOps iqdb SauceNAO cuda = 11.8.png, 70KiB, 1031x944

Anonymous Sun 21 May 2023 21:14:32 No.7388319 Report

Quoted By: >>7388327 >>7388333 >>7388336 >>7388339 >>7388348

I've been trying training at 1024 and had to adjust settings to go from batch 3 to batch 2. Does anything here looking out of place? because I got some really bad results the first time a ran this and cant see why. Its 80 images with 10 steps each.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00227-1592044654.png, 3MiB, 2048x1536

Anonymous Sun 21 May 2023 21:15:32 No.7388321 Report

Quoted By: >>7388328 >>7388563

Anonymous

View Same Google ImgOps iqdb SauceNAO 00041.png, 1MiB, 714x1000

Anonymous Sun 21 May 2023 21:16:35 No.7388322 Report

Quoted By:

Say uncle

Anonymous

Anonymous Sun 21 May 2023 21:17:28 No.7388327 Report

Quoted By: >>7388341

>>7388319
The learning rate is too high for th--
>anylora.safetensors
nvm you're just retarded

Anonymous

View Same Google ImgOps iqdb SauceNAO 00002.png, 1MiB, 768x1152

Anonymous Sun 21 May 2023 21:17:38 No.7388328 Report

Quoted By: >>7388334 >>7388343

>>7388321
why do they all have the same face

Anonymous

Anonymous Sun 21 May 2023 21:19:51 No.7388333 Report

Quoted By: >>7388341

>>7388319
> cpu threads 8
>anylora

PFFFHAHAHAHA

Anonymous

Anonymous Sun 21 May 2023 21:20:08 No.7388334 Report

Quoted By:

>>7388328
sisters

Anonymous

Anonymous Sun 21 May 2023 21:20:21 No.7388335 Report

Quoted By:

>>7388049
>>7388055
huh , the problem resolved itself seemingly out of nowhere,

weird

Anonymous

Anonymous Sun 21 May 2023 21:20:59 No.7388336 Report

Quoted By: >>7388339 >>7388341

>>7388319
>anyslop
>going from 3 to 2 instead of to 1
whats the point of this? it wont improve your results

Anonymous

Anonymous Sun 21 May 2023 21:23:42 No.7388339 Report

Quoted By: >>7388344 >>7388346

>>7388319
>>7388336
what is anylora?

Anonymous

Anonymous Sun 21 May 2023 21:23:45 No.7388340 Report

Quoted By:

>>7388008
what is my nooby self looking at here?

Anonymous

Anonymous Sun 21 May 2023 21:23:59 No.7388341 Report

Quoted By: >>7388347

>>7388327
>>7388333
>>7388336
ok guess trying anylora was a mistake, i'll go back to NAI. don't know whats the cpu threads do, 8 was the default with the script, 2 is just faster. Also looking at the lr i think i missed a zero... starting to see why it came out a jumbled mess.

Anonymous

Anonymous Sun 21 May 2023 21:24:38 No.7388343 Report

Quoted By:

>>7388328
High res fix tends to average certain aspects on the render, mainly faces, skin tones and specific signals of the prompt; best way to fight it is inpainting or blending different renders in Photoshop.
Heard there are some extensions to detect faces, but haven't explored those yet

Anonymous

Anonymous Sun 21 May 2023 21:25:20 No.7388344 Report

Quoted By:

>>7388339
meant to be a neutral thing to train on, look it up on civit
I havent actually seen anyone do x/y tests of NAI vs any so hard to tell which is better, but I'd trust NAI, obviously

Anonymous

Anonymous Sun 21 May 2023 21:26:15 No.7388346 Report

Quoted By: >>7388351 >>7388362

>>7388339
It's a discord model. They train on it because training on NAI is too "outdated" and NAI is "diluted" in newer models.

Anonymous

Anonymous Sun 21 May 2023 21:26:20 No.7388347 Report

Quoted By: >>7388350

>>7388341
>2 is just faster
yes but 3 is faster than 2 so why go down to 2?
>missed a zero
use easyscript popups dumbass and just save the json if you really want it as a text file.

Anonymous

Anonymous Sun 21 May 2023 21:26:27 No.7388348 Report

Quoted By:

>>7388319
Hard to say since you gave no clues about what is bad and what you're trying to do. Just change text encoder to 0.00002 and it should be fine (also train on NAI).

Anonymous

Anonymous Sun 21 May 2023 21:26:45 No.7388349 Report

Quoted By: >>7388375

>>7388036
save your money and buy a cheap older card.... don't feed the beast. these cloud providers can go fuck themselves for lobbying to force people into their ecosystem

Anonymous

Anonymous Sun 21 May 2023 21:28:34 No.7388350 Report

Quoted By:

>>7388347
free collab cant do 1024 batch 3

Anonymous

Anonymous Sun 21 May 2023 21:29:01 No.7388351 Report

Quoted By: >>7388355

>>7388346
>NAI is "diluted" in newer models.
then whats the model that one needs to train on?

Anonymous

Anonymous Sun 21 May 2023 21:30:39 No.7388355 Report

Quoted By: >>7388359 >>7388365 >>7388370

>>7388351
anylora, 100 repeats, 1 epoch, 15 images

Anonymous

Anonymous Sun 21 May 2023 21:31:39 No.7388358 Report

Quoted By: >>7388361 >>7388372

how tf do I update the pip in stable diffusion? It updated on windows but even on a fresh install it's saying it's out of date

Anonymous

Anonymous Sun 21 May 2023 21:31:47 No.7388359 Report

Quoted By:

>>7388355
this desu

Anonymous

Anonymous Sun 21 May 2023 21:32:35 No.7388361 Report

Quoted By: >>7388367

>>7388358
you are not updating it from the venv most likely

Anonymous

Anonymous Sun 21 May 2023 21:32:37 No.7388362 Report

Quoted By: >>7388383

>>7388346
I'm always surprised to hear how retarded they are. It doesn't matter that NAI is old, it's still the core of pretty much every model mix becasue we haven't had any better base models for anime since the leak.

Anonymous

Anonymous Sun 21 May 2023 21:33:19 No.7388365 Report

Quoted By:

>>7388355
>15 images
fuck off newfag
it's 5

Anonymous

Anonymous Sun 21 May 2023 21:34:26 No.7388367 Report

Quoted By: >>7388368 >>7388369

>>7388361
where do I do it from? when I follow the instructions it gives, I open the python exe and it just tells me the syntax is incorrect. Windows took it no problem, so I'm pretty sure I'm doing it right

Anonymous

Anonymous Sun 21 May 2023 21:35:00 No.7388368 Report

Quoted By: >>7388373

>>7388367
>I open the python exe
Anon-kun...

Anonymous

Anonymous Sun 21 May 2023 21:35:17 No.7388369 Report

Quoted By:

>>7388367
>Windows took it no problem, so I'm pretty sure I'm doing it right
pffft

Anonymous

Anonymous Sun 21 May 2023 21:35:28 No.7388370 Report

Quoted By:

>>7388355
oh anylora is a checkpoint, whats about it that makes it nuetral?
how was it trained or mixed?

Anonymous

Anonymous Sun 21 May 2023 21:35:38 No.7388371 Report

Quoted By: >>7388374

theres no issue with mixing png and jpg for training right?

Anonymous

Anonymous Sun 21 May 2023 21:35:55 No.7388372 Report

Quoted By:

>>7388358
set the PIP_DISABLE_PIP_VERSION_CHECK env var to 1 if you just want to ignore it

Anonymous

Anonymous Sun 21 May 2023 21:36:01 No.7388373 Report

Quoted By:

>>7388368
that's what the instructions say so..

Anonymous

Anonymous Sun 21 May 2023 21:36:39 No.7388374 Report

Quoted By: >>7388377

>>7388371
no, just makesure the filenames are not the same, I fucked a lora once because 1.jpg and 1.png created the same 1.txt file

Anonymous

Anonymous Sun 21 May 2023 21:36:58 No.7388375 Report

Quoted By: >>7388381

>>7388349
I would if I could, trust me, but due to life circumstances I can only use laptop for now.
I can prompt on this laptop but it's pretty slow and I don't want to fuck it up with excess proompting. Hopefully I'll be able to use a proper PC in a ~year.

Anonymous

Anonymous Sun 21 May 2023 21:37:45 No.7388376 Report

Quoted By: >>7388384

can anyone recommend a guy for a 3d model commission based off of a 2D image? it's for an oppai loli with off the charts tits and ass (not hyper tits/ass, but close) so i can use the model for training a lora

Anonymous

Anonymous Sun 21 May 2023 21:37:48 No.7388377 Report

Quoted By:

>>7388374
yea i got that covered, was upscaling jpgs which turned them to pngs, just copy pasted the names from old files

Anonymous

Anonymous Sun 21 May 2023 21:38:42 No.7388381 Report

Quoted By: >>7388387

>>7388375
invest that money into a higher AC bill then and you won't have to worry about messing it up. that, or look into eGPUs (no clue how viable they are in 2023).

Anonymous

Anonymous Sun 21 May 2023 21:38:45 No.7388382 Report

Quoted By: >>7388385

so what are the new models that nai is suppose to be diluted in?

Anonymous

Anonymous Sun 21 May 2023 21:38:56 No.7388383 Report

Quoted By: >>7388399 >>7388402 >>7388404

>>7388362
It's fried to shit btw.
>Training on this model is much more effective conpared to NAI, so at the end you might want to adjust the weight or offset (I suspect that's because NAI is now much diluted in newer models). I usually find good results at 0.65 weigth that I later offset to 1.
Basically he trains fried LoRAs then just re-saves them with reduced weight; it's hilarious.

Anonymous

Anonymous Sun 21 May 2023 21:39:23 No.7388384 Report

Quoted By: >>7388393

>>7388376
>3d model commission based off of a 2D image?
for vrchat or for 3d printing?

Anonymous

Anonymous Sun 21 May 2023 21:40:03 No.7388385 Report

Quoted By: >>7388391

>>7388382
We're making fun of discord trannies. Nobody is agreeing with them.

Anonymous

Anonymous Sun 21 May 2023 21:41:49 No.7388387 Report

Quoted By: >>7388392

>>7388381
The air temps are fine at my place but I can't monitor VRAM junction temps. I even have a cooling pad
>eGPUs
The cases cost as much as the fucking GPU in my shithole cuntry. I hate cloudshit but it's just the better option for me now.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00006-998629269-1girl, hair orna (...).png, 3MiB, 1560x2184

Anonymous Sun 21 May 2023 21:43:01 No.7388388 Report

Quoted By:

The Tsukis yearn for the mines.

Anonymous

Anonymous Sun 21 May 2023 21:47:21 No.7388391 Report

Quoted By: >>7388399

>>7388385
do you think it might be a bit shortsighted to discount things purely out of website rivalries?

Anonymous

Anonymous Sun 21 May 2023 21:47:55 No.7388392 Report

Quoted By: >>7388401

what does everyone think about this?

"I would like to give you a suggestion. With the increase of lora models, many characters have multiple lora. Generally speaking, the file size of lora is determined by the dim parameter, such as 8M 18M 36M 72M 144M and so on. The larger the lora, the better the effect when switching to a large model, especially when an anime character is switched to a live model. According to my test, if you only make animation pictures, about 36M lora is enough, if you want to make real-life pictures, at least 72M lora should be used. Of course, the lora produced by some experts can appropriately relax this range."

>>7388387
understood. well godspeed anon, hopefully you can get your new PC soon to do epic gens

Anonymous

Anonymous Sun 21 May 2023 21:49:05 No.7388393 Report

Quoted By: >>7388394 >>7388395

>>7388384
i know literally nothing about vrchat. this is a 3D model that i will pose, take screenshots of, and use said screenshots to create a lora when i don't suck as much ass

Anonymous

Anonymous Sun 21 May 2023 21:49:36 No.7388394 Report

Quoted By:

>>7388393
3D blender model or similar* fuck

Anonymous

Anonymous Sun 21 May 2023 21:49:42 No.7388395 Report

Quoted By:

>>7388393
>i know literally nothing about vrchat
why not?

Anonymous

Anonymous Sun 21 May 2023 21:50:06 No.7388396 Report

Quoted By:

>>7387927
it's a shame that the top frame of the glasses blended in with the eyes, this one is pretty sweet

Anonymous

Anonymous Sun 21 May 2023 21:51:50 No.7388399 Report

Quoted By:

>>7388391
>purely
>meanwhile the model author:
>>7388383

Anonymous

View Same Google ImgOps iqdb SauceNAO image_2023-05-22_015136457.png, 516KiB, 512x768

Anonymous Sun 21 May 2023 21:51:53 No.7388400 Report

Quoted By:

Anonymous

Anonymous Sun 21 May 2023 21:52:21 No.7388401 Report

Quoted By:

>>7388392
gibberish

Anonymous

Anonymous Sun 21 May 2023 21:53:24 No.7388402 Report

Quoted By: >>7388419 >>7388427

>>7388383
>Basically he trains fried LoRAs then just re-saves them with reduced weight; it's hilarious.
is there a reduced quality or flexibility in doing this?

Anonymous

Anonymous Sun 21 May 2023 21:55:22 No.7388404 Report

Quoted By:

>>7388383
>Training on this model is much more effective conpared to NAI, so at the end you might want to adjust the weight or offset (I suspect that's because NAI is now much diluted in newer models). I usually find good results at 0.65 weigth that I later offset to 1.
This is fucking hilarious

Anonymous

Anonymous Sun 21 May 2023 21:58:23 No.7388408 Report

Quoted By:

the guy doesn't even tell you what's in his model

Anonymous

Anonymous Sun 21 May 2023 21:59:03 No.7388409 Report

Quoted By:

after updating my xformers and pytorch i get this in the img2img tab:
RuntimeError: Input type (torch.cuda.HalfTensor) and weight type (torch.HalfTensor) should be the same
sometimes after running gens there and i can't use anything in img2img without restarting the whole webui
what the heck

Anonymous

View Same Google ImgOps iqdb SauceNAO 01174-3707367263.png, 1MiB, 1024x1536

Anonymous Sun 21 May 2023 22:01:12 No.7388411 Report

Quoted By: >>7388414 >>7388939

Anonymous

Anonymous Sun 21 May 2023 22:02:15 No.7388414 Report

Quoted By: >>7388420 >>7388429

>>7388411
gat- wait no it's that schizo name model right?

Anonymous

Anonymous Sun 21 May 2023 22:03:47 No.7388419 Report

Quoted By: >>7388427

>>7388402
I assume its the same as resizing, aka yes there is quality loss, probably more so than resize.

Anonymous

Anonymous Sun 21 May 2023 22:04:03 No.7388420 Report

Quoted By: >>7388425

>>7388414
https://files.catbox.moe/eafrdz.png

Anonymous

View Same Google ImgOps iqdb SauceNAO 00168-205377750.png, 2MiB, 1024x1536

Anonymous Sun 21 May 2023 22:06:05 No.7388424 Report

Quoted By:

okay installation unfucked but sometimes the python task will run at 8gb ram idle and idk why

Anonymous

Anonymous Sun 21 May 2023 22:06:10 No.7388425 Report

Quoted By: >>7388429

>>7388420
Oh, it's that one. I tested it already. Thanks anyway.
Is it just me or did people start posting way better gens suddenly, btw?

Anonymous

Anonymous Sun 21 May 2023 22:06:41 No.7388427 Report

Quoted By: >>7388430 >>7388454

>>7388402
It's the same as using a deepfried Civit LoRA on low weight. Try it out.
>>7388419
Nah, resizing is very different.

Anonymous

Anonymous Sun 21 May 2023 22:07:13 No.7388429 Report

Quoted By: >>7388431

>>7388414
No. They were a little similar on his other post but still quite obvious.

>>7388425
its just posters actually posting lol

>>7387962
do tell how it goes

Anonymous

Anonymous Sun 21 May 2023 22:08:38 No.7388430 Report

Quoted By:

>>7388427
Oh, its just a slider illusion where he sets 0.6 as 1?

Anonymous

Anonymous Sun 21 May 2023 22:08:50 No.7388431 Report

Quoted By:

>>7388429
>its just posters actually posting lol
fair
>do tell how it goes
Actually I suddenly want to gen, so tooomorrow????
I must postpone this as long as I can!

Anonymous

Anonymous Sun 21 May 2023 22:09:54 No.7388432 Report

Quoted By: >>7388440

I wont run into issues if I increase max tokens for training, right?

Anonymous

Anonymous Sun 21 May 2023 22:15:39 No.7388440 Report

Quoted By: >>7388444

>>7388432
It's hard to say.

Anonymous

View Same Google ImgOps iqdb SauceNAO image_2023-05-22_021452932.png, 2MiB, 1280x1920

Anonymous Sun 21 May 2023 22:15:49 No.7388442 Report

Quoted By: >>7388586 >>7388588 >>7389120

Anonymous

Anonymous Sun 21 May 2023 22:16:41 No.7388443 Report

Quoted By: >>7388450 >>7388461

>>7387962
Too lazy to dig up images but I can tell you that batch 1 is the only one that has flipped between "okay" and "unholy amalgamations of flesh" using identical settings with different (but very similar) datasets.

Anonymous

Anonymous Sun 21 May 2023 22:17:16 No.7388444 Report

Quoted By: >>7388450

>>7388440
welp, guess I'll stick to 150.
Gotta wonder which is better between 150/225 if around a quarter of the set has over 150 tokens. Hopefully it doesnt matter much

Anonymous

Anonymous Sun 21 May 2023 22:19:37 No.7388450 Report

Quoted By:

>>7388443
It's weird because theoretically you would want to train at batch one. I might just be too lazy to properly test.
>>7388444
If I had to guess, it might not be good because you'll get less dropout if you do, and that could dilute your tag strength.
But then also it might make it more responsive. Just gotta bake two.

Anonymous

View Same Google ImgOps iqdb SauceNAO Mauville.jpg, 255KiB, 1024x1536

Anonymous Sun 21 May 2023 22:21:44 No.7388452 Report

Quoted By: >>7388643

>>7387969
Inpaint until you like it.

Anonymous

Anonymous Sun 21 May 2023 22:22:49 No.7388454 Report

Quoted By: >>7388459

>>7388427
>It's the same as using a deepfried Civit LoRA on low weight.
and does that reduce quality or flexibility ?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00019-369653789-1girl, wolf girl (...).png, 3MiB, 1560x2184

Anonymous Sun 21 May 2023 22:25:37 No.7388456 Report

Quoted By:

heh

Anonymous

Anonymous Sun 21 May 2023 22:27:51 No.7388459 Report

Quoted By: >>7388464

>>7388454
It does. Character resemblance fades as you decrease the weight while trying to reduce the deepfry effect.

Anonymous

View Same Google ImgOps iqdb SauceNAO 1678597094690798-3641094407-pov, (...).png, 1MiB, 1304x976

Anonymous Sun 21 May 2023 22:28:41 No.7388461 Report

Quoted By: >>7388475

>>7388443
that doesnt sound right, I've trained on many diff batch1 datasets and while some came out worse, I did run most of those on batch2 to test if that would fix the issue but its never helped. If anything it'd have to do with learning rate/te changes than batch size.
I've also never had a dataset produce unholy shit unless I used some retarded settings, even with sets where most of it is unholy sex poses that cant possibly make sense to the AI I've only had mild issues.

My only persisting problem is backgrounds and while some of it is probably me missing some tag or a single retarded image running it, I cant help but wonder if running a batch size of 5 would have magically fixed some of these. Or should I be adjusting dim/alpha LR/TE in these cases to "underbake" it (which is probably pretty similar to "generalization" except better)?

Anonymous

View Same Google ImgOps iqdb SauceNAO file.png, 1MiB, 1024x1152

Anonymous Sun 21 May 2023 22:31:08 No.7388464 Report

Quoted By: >>7388468 >>7388482

>>7388459
thats not necessarily right - if its deepfried then it wont fade as quickly as a normal one. Theres alot of civit loras that work perfectly fine at 0.6 or so, pic rel. Only issue is if it's so fried that you cant obtain these low resemblance effects because it would continue going strong around 0.1 but idk if thats even possible

Anonymous

Anonymous Sun 21 May 2023 22:32:25 No.7388468 Report

Quoted By: >>7388476

>>7388464
I fucking knew it's you

Anonymous

Anonymous Sun 21 May 2023 22:33:26 No.7388471 Report

Quoted By: >>7388477

That neutral prompt extensions is finished, I think.

Anonymous

Anonymous Sun 21 May 2023 22:35:12 No.7388475 Report

Quoted By:

>>7388461
Yeah, I just checked and the unholy results were actually because the lora was baked with absurdly high steps for testing purposes. I went back to 2 anyway since I have a better feel for the settings with it. 1 seems to tolerate retarded settings less gracefully.

Anonymous

Anonymous Sun 21 May 2023 22:35:27 No.7388476 Report

Quoted By:

>>7388468
Im not the guy you responded to though. You should know by now I'd say it like I did with the Inori post rather than waste time leading with questions. lol

Anonymous

Anonymous Sun 21 May 2023 22:35:40 No.7388477 Report

Quoted By: >>7388486

>>7388471
So what we do with it now?

Anonymous

Anonymous Sun 21 May 2023 22:40:27 No.7388481 Report

Quoted By:

>>7388289
apologies for lashing out, I don't agree with you but arguing about this is pointless

Anonymous

Anonymous Sun 21 May 2023 22:41:02 No.7388482 Report

Quoted By: >>7388490

>>7388464
Fried loras are often workable if you really want to and better than nothing but generally are shittier than unfried ones, especially when you combine them
There are of course unsalvageable ones with 5-10 pic datasets at 100 repeats

Anonymous

Anonymous Sun 21 May 2023 22:42:50 No.7388484 Report

Quoted By: >>7388498

>>7388268
I finally get it. gomennasorry...

Anonymous

Anonymous Sun 21 May 2023 22:43:31 No.7388486 Report

Quoted By: >>7389456

>>7388477
easier to prompt contradictory concepts
easier to use overlapping tokens in the positive and negative prompt (I think)

Anonymous

Anonymous Sun 21 May 2023 22:46:01 No.7388490 Report

Quoted By: >>7388716

>>7388482
The ZnT Louise ones come to mind and some unholy Kurisu one that I couldn't get to work for me (might be skill issue since it seemed to work really well when/if it worked).
Though it varies, some of them are fried nicely like the Inori - you either get 1:1 anime outputs that you can tard wrangle into something decent, or a workable lora at 0.5.

Anonymous

Anonymous Sun 21 May 2023 22:57:20 No.7388498 Report

Quoted By:

>>7388484
heh

Anonymous

Anonymous Sun 21 May 2023 23:06:49 No.7388510 Report

Quoted By: >>7388512 >>7388516

I'm asking over in /g/ as well, but I thought someone here might know. Does anyone know how to use Regional Prompter with multiple Loras?

I'm trying to do this girl on the left
https://civitai.com/models/43696/shuten-douji-fate-grand-order

and this girl on the right
https://civitai.com/models/11798?modelVersionId=13941

I do have Composable Lora set up.

Anonymous

Anonymous Sun 21 May 2023 23:10:27 No.7388512 Report

Quoted By: >>7388515

>>7388510
>Composable Lora
that's for latent couple

Anonymous

View Same Google ImgOps iqdb SauceNAO 00014-1252838234.png, 2MiB, 1024x1536

Anonymous Sun 21 May 2023 23:11:08 No.7388513 Report

Quoted By: >>7388523 >>7388529 >>7388591

so I've settle on a workflow that allows me to train a lora for a character in a particular art style while maintaining a decent amount of flexibility, anyone loras that feel too rigid which could use retraining?

Anonymous

Anonymous Sun 21 May 2023 23:11:58 No.7388515 Report

Quoted By:

>>7388512
So Regional Prompter can't do two Loras at all?

Anonymous

Anonymous Sun 21 May 2023 23:12:33 No.7388516 Report

Quoted By:

>>7388510
use 'Latent' for generation mode, place lora in correct part of the prompt.

Anonymous

Anonymous Sun 21 May 2023 23:15:50 No.7388523 Report

Quoted By:

>>7388513
Share your method anon, sounds like it would make a useful reference.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00022-603421652.png, 1MiB, 1024x1536

Anonymous Sun 21 May 2023 23:18:19 No.7388529 Report

Quoted By: >>7388591

>>7388513

Anonymous

Anonymous Sun 21 May 2023 23:30:39 No.7388551 Report

Quoted By: >>7388608

>>7388275
>>7388298
Catbox please?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00059-3233087806.png, 4MiB, 2136x1152

Anonymous Sun 21 May 2023 23:36:07 No.7388563 Report

Quoted By:

>>7388321
inspiring

Anonymous

View Same Google ImgOps iqdb SauceNAO 00094-3110614712-nsfw, scan, ((t (...).png, 2MiB, 1024x1536

Anonymous Sun 21 May 2023 23:36:11 No.7388564 Report

Quoted By: >>7388574 >>7388578

Anonymous

Anonymous Sun 21 May 2023 23:43:37 No.7388574 Report

Quoted By:

>>7388564
the loli philanderer returns ;)

Anonymous

Anonymous Sun 21 May 2023 23:46:10 No.7388578 Report

Quoted By:

>>7388564
sexo.

Anonymous

Anonymous Sun 21 May 2023 23:46:54 No.7388580 Report

Quoted By:

>>7387801
catbox?

Anonymous

View Same Google ImgOps iqdb SauceNAO 06964-2679512510-,(masterpiece,h (...).png, 1MiB, 1024x1536

Anonymous Sun 21 May 2023 23:49:01 No.7388584 Report

Quoted By: >>7388595

Anonymous

Anonymous Sun 21 May 2023 23:52:36 No.7388586 Report

Quoted By:

>>7388442
BASED

Anonymous

Anonymous Sun 21 May 2023 23:54:54 No.7388588 Report

Quoted By: >>7388602

>>7388442
got catbox?

Anonymous

Anonymous Sun 21 May 2023 23:57:30 No.7388591 Report

Quoted By:

>>7388513
>>7388529
Is that louise? If so mind if you share the LoRA please?

Anonymous

Anonymous Mon 22 May 2023 00:01:03 No.7388595 Report

Quoted By: >>7388613

>>7388584
catbox s.v.p.

Anonymous

Anonymous Mon 22 May 2023 00:06:24 No.7388602 Report

Quoted By: >>7388703

>>7388588
https://files.catbox.moe/ro7j80.png

Anonymous

Anonymous Mon 22 May 2023 00:07:36 No.7388606 Report

Quoted By: >>7388635

anyone have trouble with ram usage? a week ago the webui would use maybe 2.5 gb of ram idle but now it uses 7 gb and sometimes won't let me gen if i so much as watch a video in chrome

Anonymous

View Same Google ImgOps iqdb SauceNAO catbox_zq0vqw.png, 2MiB, 1280x1536

Anonymous Mon 22 May 2023 00:08:05 No.7388608 Report

Quoted By:

>>7388551
https://files.catbox.moe/bq7z2r.png

Anonymous

Anonymous Mon 22 May 2023 00:09:23 No.7388609 Report

Quoted By: >>7388612 >>7388615 >>7388616 >>7388900

Is there any hopium for the future here? I hate to admit it, but I can identify AI art just from a thumbnail, all the mixes have the same giveaways. Are there any major developments towards improving the technology or are we just going to be recycling the same mixes for years?

Anonymous

Anonymous Mon 22 May 2023 00:10:33 No.7388612 Report

Quoted By:

>>7388609
dude a lot of a shit happened in just 5 months,
dont go all doom when things slow down.

Anonymous

View Same Google ImgOps iqdb SauceNAO 06965-647878519-,(masterpiece,hi (...).png, 2MiB, 1024x1536

Anonymous Mon 22 May 2023 00:10:38 No.7388613 Report

Quoted By:

>>7388595
https://files.catbox.moe/i87bqh.png

Anonymous

Anonymous Mon 22 May 2023 00:11:21 No.7388615 Report

Quoted By:

>>7388609
i can conjure pictures of curvy yordle girls taking dick so i'm content, anything that can make it easier is just a bonus at this point

Anonymous

Anonymous Mon 22 May 2023 00:11:54 No.7388616 Report

Quoted By:

>>7388609
>but I can identify AI art just from a thumbnail
maybey but a lot of things get caught in the crossfire

Anonymous

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

Anonymous Mon 22 May 2023 00:13:04 No.7388619 Report

Quoted By: >>7388628

Finally got something acceptable

Anonymous

View Same Google ImgOps iqdb SauceNAO file.png, 2MiB, 1536x1536

Anonymous Mon 22 May 2023 00:19:26 No.7388628 Report

Quoted By:

>>7388619
Testing with higher res

Anonymous

Anonymous Mon 22 May 2023 00:28:05 No.7388635 Report

Quoted By:

>>7388606
i dont know what i did but apparently i installed webui twice and practically every file has a duplicate with a (2), maybe this is why? maybe it's running two consecutive webuis in that single python task? doesnt make sense to me but i'll see after i clean it up

Anonymous

Anonymous Mon 22 May 2023 00:30:49 No.7388637 Report

Quoted By: >>7388645

so im told i should be upset that someone made money off my loras

Anonymous

View Same Google ImgOps iqdb SauceNAO catbox_wpm20m.png, 1MiB, 1024x1280

Anonymous Mon 22 May 2023 00:35:23 No.7388640 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO Nightcap.jpg, 176KiB, 1024x1536

Anonymous Mon 22 May 2023 00:38:02 No.7388643 Report

Quoted By:

>>7388452
Controlnet until you nede it.

Anonymous

Anonymous Mon 22 May 2023 00:38:28 No.7388645 Report

Quoted By:

>>7388637
it wasn't me, officer

Anonymous

View Same Google ImgOps iqdb SauceNAO 00031-2104361576-1girl, official (...).png, 2MiB, 1560x2184

Anonymous Mon 22 May 2023 00:46:17 No.7388652 Report

Quoted By:

Anonymous

Anonymous Mon 22 May 2023 00:52:00 No.7388664 Report

Quoted By:

>switch back to torch==0.13.1 torchvision==0.14.1 xformers==0.0.16
>it/s increased by 1
lol ok I'll cope with non determinism

Anonymous

Anonymous Mon 22 May 2023 00:52:06 No.7388665 Report

Quoted By:

I clench my fist and beat it I bust my nut and freeze it, I light my stove and burn it, I am the jizz I worship

Anonymous

View Same Google ImgOps iqdb SauceNAO 1657697164181518.jpg, 291KiB, 1408x1792

Anonymous Mon 22 May 2023 00:56:19 No.7388671 Report

Quoted By: >>7388673 >>7388766

Anonymous

Anonymous Mon 22 May 2023 00:57:14 No.7388673 Report

Quoted By:

>>7388671
gato

Anonymous

View Same Google ImgOps iqdb SauceNAO 24046-1085514488-1girl, breasts, (...).png, 1MiB, 1536x1024

Anonymous Mon 22 May 2023 01:15:38 No.7388687 Report

Quoted By:

>cells at work black is weirdly good at tentacle rape
interdasting

Anonymous

View Same Google ImgOps iqdb SauceNAO 00082-1523695520-1girl, hair orn (...).png, 3MiB, 1560x2184

Anonymous Mon 22 May 2023 01:23:00 No.7388694 Report

Quoted By:

They always make those female armors so impractical!

Anonymous

View Same Google ImgOps iqdb SauceNAO 00002-1713341183.png, 2MiB, 1024x1536

Anonymous Mon 22 May 2023 01:25:22 No.7388698 Report

Quoted By:

Anonymous

Anonymous Mon 22 May 2023 01:34:48 No.7388703 Report

Quoted By: >>7388751

>>7388602
>opllv2e5
What is this?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00094-3026188779-1girl, mature f (...).png, 4MiB, 1560x2184

Anonymous Mon 22 May 2023 01:37:54 No.7388708 Report

Quoted By:

>"Trust me, bub. I will be worth the Chilli's gift card and three Dr. Peppers"

Anonymous

Anonymous Mon 22 May 2023 01:51:53 No.7388716 Report

Quoted By: >>7389190

>>7388490
>ZnT Louise
>Kurisu
had problem with both of these actually
The ultimate fry-test is using the loras on models that are very far from NAI (like furry models). Loras I suspected were shit all turned out to completely fry the output before actually applying the concept.
The ones I made mostly work fine

Anonymous

Anonymous Mon 22 May 2023 02:01:46 No.7388730 Report

Quoted By: >>7388733 >>7388735 >>7388737

Out of curiosity, what prompts do you think are still the most difficult to generate accurately? From what I've seen, I'd say anything involving smoking still produces weird shit most of the time.

Anonymous

Anonymous Mon 22 May 2023 02:07:12 No.7388733 Report

Quoted By:

>>7388730
anything involving poses more complicated than standing or sitting or lying
same with sex positions

Anonymous

Anonymous Mon 22 May 2023 02:08:22 No.7388735 Report

Quoted By:

>>7388730
holding sword

Anonymous

Anonymous Mon 22 May 2023 02:09:31 No.7388737 Report

Quoted By:

>>7388730
more than 1 person, interacting, and they're existing characters

Anonymous

Anonymous Mon 22 May 2023 02:22:49 No.7388751 Report

Quoted By: >>7388776

>>7388703
opaloli lora, search hiking on kemono

Anonymous

View Same Google ImgOps iqdb SauceNAO 00238-116431471.png, 285KiB, 512x512

Anonymous Mon 22 May 2023 02:24:45 No.7388753 Report

Quoted By: >>7388755 >>7388756 >>7390664

Why does this keep happening? How do I fix this mess?

Anonymous

Anonymous Mon 22 May 2023 02:25:21 No.7388755 Report

Quoted By:

>>7388753
stop using nai

Anonymous

Anonymous Mon 22 May 2023 02:25:51 No.7388756 Report

Quoted By: >>7388757

>>7388753
use an actual porn model, not that shit

Anonymous

Anonymous Mon 22 May 2023 02:26:54 No.7388757 Report

Quoted By:

>>7388756
what would you suggest?

Anonymous

Anonymous Mon 22 May 2023 02:37:07 No.7388766 Report

Quoted By: >>7388788

>>7388671
this looks really good, what model or style is it?

Anonymous

Anonymous Mon 22 May 2023 02:42:08 No.7388769 Report

Quoted By: >>7388778 >>7388797 >>7389292

What are the current "best" models for generating fap material? i quitted on january because this stuff was destroying my dick but i kinda wanna get back to it. I recall that AOM2 was quite popular but i imagine thats not the best one anymore

Anonymous

Anonymous Mon 22 May 2023 02:48:41 No.7388776 Report

Quoted By: >>7388779

>>7388751
>+Personality excretion
really?

Anonymous

Anonymous Mon 22 May 2023 02:50:43 No.7388778 Report

Quoted By:

>>7388769
based64 i think

Anonymous

Anonymous Mon 22 May 2023 02:51:13 No.7388779 Report

Quoted By:

>>7388776
I don't fucking know how the two loras ended up together, the guy doesn't even post that stuff.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00300-3835172580.png, 3MiB, 1280x1920

Anonymous Mon 22 May 2023 02:55:15 No.7388784 Report

Quoted By:

Anonymous

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

Anonymous Mon 22 May 2023 02:57:14 No.7388786 Report

Quoted By: >>7388840

Anonymous

Anonymous Mon 22 May 2023 02:59:13 No.7388788 Report

Quoted By:

>>7388766
Y'shtola lora I made with a shitty blend of some models with based 64. (traditional media:1.3) does a lot of the heavy lifting.

Anonymous

Anonymous Mon 22 May 2023 03:01:17 No.7388791 Report

Quoted By: >>7388794 >>7388800 >>7388823

>>7387173
>>7387190
Thanks for proving my points faggots. Can't wait when all this AI shit gets banned everywhere

Anonymous

Anonymous Mon 22 May 2023 03:05:06 No.7388794 Report

Quoted By:

>>7388791
what are you talking about? this place is about as helpful as any other general in all of 4chan, it has nothing to do with AI per se.
Also you're not entitled to anything, and it's not like people didn't answer out of spite towards you specifically, get your head out of your ass and google stuff or hell just ask again, who gives a shit.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00745-2517443714.png, 2MiB, 1024x1536

Anonymous Mon 22 May 2023 03:05:46 No.7388797 Report

Quoted By: >>7388804

>>7388769
There are so many options now that it is almost overwhelming, and then you fall down the model mixing rabbit hole.
AOM2 is still a pretty popular choice, especially when it comes to model mixing.
AOM3 seems to have been completely forgotten.

Anonymous

Anonymous Mon 22 May 2023 03:07:15 No.7388800 Report

Quoted By:

>>7388791
Literally what was ignored

Anonymous

Anonymous Mon 22 May 2023 03:07:58 No.7388804 Report

Quoted By: >>7388806

>>7388797
>forgotten
why is that?

Anonymous

Anonymous Mon 22 May 2023 03:09:00 No.7388806 Report

Quoted By: >>7388808

>>7388804
because its shit

Anonymous

Anonymous Mon 22 May 2023 03:09:44 No.7388808 Report

Quoted By:

>>7388806
so i shoudl use AOM2nsfw then.

Anonymous

View Same Google ImgOps iqdb SauceNAO training parameters.png, 118KiB, 1327x1658

Anonymous Mon 22 May 2023 03:10:53 No.7388809 Report

Quoted By: >>7388814 >>7388815 >>7388816 >>7388821 >>7388822 >>7388854

What is the current LoRa training parameters meta? I have been way 3 weeks. I used this pic related but wasn't happy with the results. DAdaptation is not a clear winner over Adam8Bit when you want to change clothes.

Anonymous

Anonymous Mon 22 May 2023 03:22:18 No.7388814 Report

Quoted By: >>7388905

>>7388809
15 images, 100 repeats, 1 epoch, train on Anylora

Anonymous

Anonymous Mon 22 May 2023 03:23:21 No.7388815 Report

Quoted By:

>>7388809
>cosine with restarts
>on dadaptation
>No restarts

Son...

Anonymous

Anonymous Mon 22 May 2023 03:24:29 No.7388816 Report

Quoted By:

>>7388809
>that ui
lol
lmao
haha

Anonymous

Anonymous Mon 22 May 2023 03:31:53 No.7388821 Report

Quoted By:

>>7388809
Don't bother; nobody will help you. Just find a good recent LoRA and copy its settings. Nobody here wants to get into the nth training parameters argument with baiters.

Anonymous

Anonymous Mon 22 May 2023 03:32:35 No.7388822 Report

Quoted By: >>7388825 >>7388852

>>7388809
dadaptation has nothing to do with being able to change clothes if your tags are good enough, also, why are you not using the decoupling/decay arguments for it? No wonder its fucked.

Why are you shuffling the tags?

Anonymous

Anonymous Mon 22 May 2023 03:34:15 No.7388823 Report

Quoted By:

>>7388791
Lol what a pathetic crybaby faggot

Anonymous

View Same Google ImgOps iqdb SauceNAO 00017-186869439.png, 4MiB, 1720x1720

Anonymous Mon 22 May 2023 03:36:57 No.7388825 Report

Quoted By:

>>7388822

I am not the source of this parameters, an anon on here was.

>Pic related, Aqua by anon also posted on CivitAI

Anonymous

View Same Google ImgOps iqdb SauceNAO 00114-3146828480.png, 2MiB, 1536x1024

Anonymous Mon 22 May 2023 03:38:39 No.7388826 Report

Quoted By: >>7388829 >>7388897

Anonymous

View Same Google ImgOps iqdb SauceNAO 00205-3282718863.png, 538KiB, 512x768

Anonymous Mon 22 May 2023 03:39:41 No.7388829 Report

Quoted By: >>7388861

>>7388826
That's one nice ass

Anonymous

View Same Google ImgOps iqdb SauceNAO toolazytofix.png, 2MiB, 1600x1080

Anonymous Mon 22 May 2023 03:51:09 No.7388837 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO tmp_uxg0fcr.png, 900KiB, 1024x1024

Anonymous Mon 22 May 2023 03:56:29 No.7388840 Report

Quoted By:

>>7388786

Anonymous

Anonymous Mon 22 May 2023 04:24:16 No.7388852 Report

Quoted By: >>7388872

>>7388822
>Why are you shuffling the tags?
why are you not shuffling the tags? I'd dismiss it as bait normally, but the first half's reasonable.

Anonymous

Anonymous Mon 22 May 2023 04:26:02 No.7388854 Report

Quoted By:

>>7388809
1e-4
5e-5
batch 1
768
128/128

Anonymous

View Same Google ImgOps iqdb SauceNAO 00044-3532573048.png, 2MiB, 1536x1024

Anonymous Mon 22 May 2023 04:33:50 No.7388861 Report

Quoted By: >>7388897 >>7388915

>>7388829
Sadly the same level of thiccness couldn't be replicated for this gen.

Anonymous

Anonymous Mon 22 May 2023 04:45:59 No.7388872 Report

Quoted By:

>>7388852
They like dropping half their tags.

Anonymous

Anonymous Mon 22 May 2023 04:49:41 No.7388878 Report

Quoted By:

is there a "drugged" lora for character expressions?

Anonymous

Anonymous Mon 22 May 2023 05:19:48 No.7388897 Report

Quoted By: >>7388907

>>7388826
>>7388861
Catbox?

Anonymous

Anonymous Mon 22 May 2023 05:28:45 No.7388900 Report

Quoted By:

>>7388609
Have you looked at at the artists that are not absolute 0.001% best? I recently visited danbooru to get some inspiration and man most arts are certainly fucking worse than AI. Even AI hands are better than most of the shit you see. I even had to recheck the date on some of them and yeah it was made before autumn 2022.

Anonymous

Anonymous Mon 22 May 2023 05:39:16 No.7388905 Report

Quoted By: >>7388906 >>7388911

>>7388814

This actually worked? I will test further and compare.

Anonymous

Anonymous Mon 22 May 2023 05:39:57 No.7388906 Report

Quoted By: >>7388911

>>7388905
The more repeats, the better lora basically. Takes a bit more time but for me it's worth it.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00190-4141598015.png, 2MiB, 1024x1536

Anonymous Mon 22 May 2023 05:41:07 No.7388907 Report

Quoted By:

>>7388897
https://files.catbox.moe/hh5mc4.png
Inpainted the male, some of the background and the face.
https://files.catbox.moe/blmh2p.png
I photoshopped in 3 of the missing legs. Then I used img2img and inpaint to blend them in.

Anonymous

Anonymous Mon 22 May 2023 05:48:09 No.7388911 Report

Quoted By:

>>7388906
This.

>>7388905
I personally stop at 150, would go higher if it didn't take as long.

Anonymous

Anonymous Mon 22 May 2023 05:52:16 No.7388915 Report

Quoted By: >>7388927

>>7388861
catbox?

Anonymous

Anonymous Mon 22 May 2023 05:57:59 No.7388920 Report

Quoted By: >>7388925 >>7388936 >>7389290

Why are there so many bad hands embeddings? which one is the best? I heard some of them fucks with the art style.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00017-940029179.png, 1MiB, 1024x1024

Anonymous Mon 22 May 2023 05:58:20 No.7388921 Report

Quoted By:

Anonymous

Anonymous Mon 22 May 2023 05:58:42 No.7388922 Report

Quoted By:

Is civitai the only place people (read: Japanese) post loras on?
Found a lora I want on pixiv but I can't find it on civitai

Anonymous

Anonymous Mon 22 May 2023 05:59:12 No.7388925 Report

Quoted By:

>>7388920
none of them work

Anonymous

Anonymous Mon 22 May 2023 05:59:30 No.7388927 Report

Quoted By:

>>7388915
https://files.catbox.moe/2v8bb9.png
kek alright. A lot of inpainting for this one.

Anonymous

Anonymous Mon 22 May 2023 06:06:17 No.7388936 Report

Quoted By:

>>7388920
Use all of them at once till you convince yourself they work

Anonymous

Anonymous Mon 22 May 2023 06:08:00 No.7388939 Report

Quoted By:

>>7388203
>>7388263
>>7388411
Catbox please!

Anonymous

Anonymous Mon 22 May 2023 06:08:21 No.7388940 Report

Quoted By: >>7388950 >>7388969 >>7389501

hey all,
I haven't talked in this thread for a while, been more focused on other things. but I'm here with some huge news. I finally got around to making that UI that I wanted to make since I made the popup scripts.
https://github.com/derrian-distro/LoRA_Easy_Training_Scripts

This does mean that all json files are now invalid due to the change of the saving and loading system to toml files. I did have to cut a few features for v1, queues, and block weight training, but they will be added back in once I figure out how they fit into the UI. I hope everybody enjoys using the new UI, and the design choices I've made don't cause people to be annoyed.

the readme has a list of the features, and a showcase of the UI, the old scripts still exist, they just are on a branch I created for them. the old scripts might get updated every so often, but the main bulk of my work is now going to be on the UI version.

Anonymous

Anonymous Mon 22 May 2023 06:14:57 No.7388950 Report

Quoted By: >>7388979

>>7388940

Nice. . . I'll try yours since you're first few people who shared the script with us.

Anonymous

Anonymous Mon 22 May 2023 06:27:46 No.7388969 Report

Quoted By: >>7388979

>>7388940
probably still won't ever use it but looks cool, good work

Anonymous

Anonymous Mon 22 May 2023 06:33:15 No.7388979 Report

Quoted By:

>>7388950
hopefully you end up liking it, I put a lot of work into the UI.

>>7388969
that is fair, it's not surprising that not everyone will use it. Thanks for the feedback on the design though!

Anonymous

Anonymous Mon 22 May 2023 06:36:44 No.7388982 Report

Quoted By: >>7388997

Why is Desna so based bros
>Only god to actually reform the hyper evil unlike Iomedae or the cringe Gold Dragons
>Gave you prime succubus puss thats devoted to you

Anonymous

Anonymous Mon 22 May 2023 06:44:55 No.7388997 Report

Quoted By:

>>7388982
For me, it's strange beast-sama. The strongest archon

Anonymous

Anonymous Mon 22 May 2023 06:48:43 No.7389001 Report

Quoted By:

Ayy, there was an update to adetailer that fixed the blurry face when using controlnet. >>7387973

Feels good to have a hobby with a community that acts fast.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00011-4294374686.png, 2MiB, 1024x2048

Anonymous Mon 22 May 2023 07:13:33 No.7389028 Report

Quoted By: >>7389031 >>7389056

Anonymous

View Same Google ImgOps iqdb SauceNAO 00016-1980633073.png, 1MiB, 1024x1024

Anonymous Mon 22 May 2023 07:15:04 No.7389030 Report

Quoted By: >>7389034

Anonymous

Anonymous Mon 22 May 2023 07:15:15 No.7389031 Report

Quoted By: >>7389037

>>7389028
where can one get that raphtalia lora?

Anonymous

Anonymous Mon 22 May 2023 07:16:16 No.7389034 Report

Quoted By:

>>7389030
catbox?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00084-4265165755.png, 846KiB, 640x960

Anonymous Mon 22 May 2023 07:17:00 No.7389035 Report

Quoted By: >>7389036 >>7389038 >>7389056

Anonymous

Anonymous Mon 22 May 2023 07:18:20 No.7389036 Report

Quoted By: >>7389475

>>7389035
What Model/Lora?

Anonymous

Anonymous Mon 22 May 2023 07:18:24 No.7389037 Report

Quoted By: >>7389046

>>7389031
You can get it on civitai tho its kinda fucked, you can't really change the breast size without lowering the weight and putting on a big breast LoRA

Anonymous

Anonymous Mon 22 May 2023 07:18:29 No.7389038 Report

Quoted By: >>7389475

>>7389035
elfbros not like this

Anonymous

Anonymous Mon 22 May 2023 07:23:43 No.7389046 Report

Quoted By:

>>7389037
Of course it had to be fucked... Thanks anon, i appreciate the help.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00038-3243344615.png, 428KiB, 512x512

Anonymous Mon 22 May 2023 07:25:59 No.7389049 Report

Quoted By: >>7389065 >>7389068

The Thing but it's a slut.

Anonymous

View Same Google ImgOps iqdb SauceNAO 07513-2485048895-nsfw, masterpie (...).png, 988KiB, 768x1152

Anonymous Mon 22 May 2023 07:35:51 No.7389055 Report

Quoted By: >>7389064

its going to be a hot summer, I hope you have AC

Anonymous

Anonymous Mon 22 May 2023 07:36:50 No.7389056 Report

Quoted By: >>7389069 >>7389475

>>7389028
>>7389035
catbox please?

Anonymous

Anonymous Mon 22 May 2023 07:51:06 No.7389064 Report

Quoted By: >>7389988

>>7389055
Got a catbox for that?

Anonymous

Anonymous Mon 22 May 2023 07:53:39 No.7389065 Report

Quoted By:

>>7389049
Hot

Anonymous

Anonymous Mon 22 May 2023 07:56:07 No.7389068 Report

Quoted By:

>>7389049
OH MY FUCKING GOD CATBOX PLEEEASE

Anonymous

Anonymous Mon 22 May 2023 07:56:14 No.7389069 Report

Quoted By:

>>7389056
Raphtalia anon here
https://files.catbox.moe/srtjkx.png

Anonymous

View Same Google ImgOps iqdb SauceNAO ChunLiAss.png, 2MiB, 1024x1451

Anonymous Mon 22 May 2023 07:56:59 No.7389070 Report

Quoted By:

Coom...li...

Anonymous

Anonymous Mon 22 May 2023 08:03:30 No.7389073 Report

Quoted By: >>7389082 >>7389112

I am trying to get it work with my AMD card and it has been stuck at

Applying cross attention optimization (InvokeAI).
Textual inversion embeddings loaded(0):
Model loaded in 130.9s (calculate hash: 21.1s, load weights from disk: 3.9s, create model: 87.4s, apply weights to model: 6.5s, apply half(): 6.9s, load VAE: 0.6s, move model to device: 4.3s).

for like an hour or two now. Is it supposed to take this long?

Anonymous

Anonymous Mon 22 May 2023 08:10:23 No.7389082 Report

Quoted By: >>7389096 >>7389098

>>7389073
If ur using AMD ur better off using free collab

Anonymous

Anonymous Mon 22 May 2023 08:16:34 No.7389084 Report

Quoted By:

Anons are there any cool new extensions?
Also are there any other QoL extensions like the tag completion one and sd-webui-ar?

Anonymous

View Same Google ImgOps iqdb SauceNAO 20264-lp_ai_img.png, 2MiB, 1152x1728

Anonymous Mon 22 May 2023 08:23:06 No.7389094 Report

Quoted By:

Anonymous

Anonymous Mon 22 May 2023 08:25:08 No.7389096 Report

Quoted By: >>7389102

>>7389082
free colab is dead

Anonymous

Anonymous Mon 22 May 2023 08:25:51 No.7389098 Report

Quoted By:

>>7389082
I got it working! I just closed it and reopened it and it worked!

Anonymous

Anonymous Mon 22 May 2023 08:34:10 No.7389102 Report

Quoted By: >>7389103

>>7389096
Paperspace was shut down?

Anonymous

Anonymous Mon 22 May 2023 08:36:43 No.7389103 Report

Quoted By: >>7389113

>>7389102
Paperspace is not colab
And free paperspace is barely usable

Anonymous

Anonymous Mon 22 May 2023 08:42:34 No.7389109 Report

Quoted By: >>7389111 >>7389338

Does anyone here add the description/screenshot tags to the lora with the additional networks extension? I am considering doing this and wondering if this is too autistic or not

Anonymous

Anonymous Mon 22 May 2023 08:45:22 No.7389111 Report

Quoted By: >>7389114

>>7389109
I don't. Seems counterproductive.

Anonymous

Anonymous Mon 22 May 2023 08:48:54 No.7389112 Report

Quoted By:

>>7389073
Are you using the wiki guide?

Anonymous

Anonymous Mon 22 May 2023 08:50:26 No.7389113 Report

Quoted By:

>>7389103
Pretty sure it's still called colab.. hm

Anonymous

Anonymous Mon 22 May 2023 08:51:36 No.7389114 Report

Quoted By: >>7389131

>>7389111
>counterproductive
Why? Would help with tagging and seeing an example. Some loras dont have training info on them also

Anonymous

Anonymous Mon 22 May 2023 08:53:17 No.7389117 Report

Quoted By: >>7389118

>>7388071
too lazy to read this so someone confirm chatgpt did a decent job translating this repetitive blogpost:

> User trained a language model (LoRA) to generate "waifu" images.
> Used a dataset of 25 safe-for-work waifu artworks, ignoring recommended image quantity.
> Kept original tags but removed artist tags, applied same tags to different images.
> Trained the model up to epoch 10 and achieved good results with a specific prompt.
> Retrained the model after gathering a larger dataset of 427 images, filtering unwanted tags, and organizing the dataset properly.
> Results with the retrained model were not as good due to insufficient training.
> Sought advice from a forum and retrained the model using a dataset of 72 similar curated images.
> Removed artists' signatures and some background elements to reduce noise in the dataset.
> Organized the dataset and trained the model up to epoch 10.
> Results with the final model were inconsistent, with some epochs undercooked and others overcooked.
> Desired sketch style was missing in the generated images.
> Tested the initial model with the prompt "embarrassment," but the generated image didn't capture the intended emotions.
> Overall, the user faced mixed results and encountered difficulties in achieving the desired output with the trained models.

Anonymous

Anonymous Mon 22 May 2023 08:54:14 No.7389118 Report

Quoted By: >>7389119

>>7389117
> User fucked around and found out

Anonymous

Anonymous Mon 22 May 2023 08:57:09 No.7389119 Report

Quoted By:

>>7389118
>In post-nut clarity, the user decided to retrain the model. They gathered a larger dataset of 427 images, filtering out unwanted tags and organizing the dataset properly. However, when they loaded the model trained on this dataset and used the same prompt, the results were not as good due to insufficient training.
kek i'm disappointed it didn't carry this over when translating it to bullet points

Anonymous

Anonymous Mon 22 May 2023 08:57:19 No.7389120 Report

Quoted By: >>7389124

>>7388442
God DAMNNNN

Anonymous

Anonymous Mon 22 May 2023 08:58:00 No.7389124 Report

Quoted By:

>>7389120
shut the fuck up

Anonymous

Anonymous Mon 22 May 2023 09:07:25 No.7389128 Report

Quoted By: >>7389131

How the fuck do I inpaint an object in hand?

Anonymous

Anonymous Mon 22 May 2023 09:09:16 No.7389131 Report

Quoted By: >>7389133

>>7389114
Because the end result all depends on a myriad of other things.
Also I'm schizo so I'd be going back and reuploading them over and over because I found a better image.
>>7389128
Paint it on in MS Paint and go over it with inpaint.

Anonymous

Anonymous Mon 22 May 2023 09:12:50 No.7389132 Report

Quoted By:

>>7386260
catbox?

Anonymous

Anonymous Mon 22 May 2023 09:15:16 No.7389133 Report

Quoted By: >>7389135 >>7389136

>>7389131
what settings should be used? some anons are doing miracles with their inpaints yet I'm getting some weird abominations

Anonymous

Anonymous Mon 22 May 2023 09:17:46 No.7389135 Report

Quoted By:

>>7389133
they'll come to you in a dream, just slumber and wait

Anonymous

Anonymous Mon 22 May 2023 09:18:48 No.7389136 Report

Quoted By: >>7389139 >>7389142

>>7389133
This is what I use. >>7385398

Anonymous

Anonymous Mon 22 May 2023 09:23:30 No.7389139 Report

Quoted By: >>7389145

>>7389136 me
>try to get a sword to inpaint as an example because some other anon said it's impossible to gen
>actually get a good sword holding pic
lmao

Anonymous

Anonymous Mon 22 May 2023 09:25:11 No.7389142 Report

Quoted By: >>7389147 >>7389149

>>7389136

I'll try it when I get around, thanks

So people delete their entire prompt if they use only masked and keep it if using whole picture?

Anonymous

Anonymous Mon 22 May 2023 09:26:14 No.7389145 Report

Quoted By:

>>7389139
It's not impossible but it's too gacha. I used inpaint for swords too

Anonymous

Anonymous Mon 22 May 2023 09:26:49 No.7389147 Report

Quoted By:

>>7389142
Yeah.
You can still only add to the existing prompt on only masked if you do low enough denoise, but otherwise you do need to only specify what you're inpainting if using it.

Anonymous

Anonymous Mon 22 May 2023 09:27:35 No.7389149 Report

Quoted By: >>7389301

>>7389142
the prompt matters less than you think it would. Well, usually even with "only masked" you can (should?) max out padding and then you dont even have to type anything if its "obvious" - it knows there should be a finger/hand there for example.
Or you can still leave the whole prompt in, its up to you.
This changes when you reduce pixel padding to a low number cause it doesnt have image context if you do so

Anonymous

Anonymous Mon 22 May 2023 10:01:14 No.7389190 Report

Quoted By:

>>7388716
>The ultimate fry-test is using the loras on models that are very far from NAI (like furry models). Loras I suspected were shit all turned out to completely fry the output before actually applying the concept.
Interesting
I tested NAI-trained loras on the vtuber models and they seemed to be actually weaker and didn't really fry much

Anonymous

Anonymous Mon 22 May 2023 10:02:02 No.7389193 Report

Quoted By:

>>7388010
not that anon but care to explain how to do that, preferebly in a screencap or two.
I've got controlnet and multidiffusion

Anonymous

View Same Google ImgOps iqdb SauceNAO 00069-1633342976.png, 1MiB, 1280x1024

Anonymous Mon 22 May 2023 10:39:17 No.7389228 Report

Quoted By:

Anonymous

Anonymous Mon 22 May 2023 10:43:50 No.7389234 Report

Quoted By: >>7389311

this lykon civit guy is almost as bad as that 100 repeat guidemaker, jesus

Anonymous

Anonymous Mon 22 May 2023 10:44:59 No.7389235 Report

Quoted By:

Any way for colab lora trainers to resize dim? I really don't want to install all the kohya/easyscript shit just to resize dim

Anonymous

View Same Google ImgOps iqdb SauceNAO 00276-1719150929.png, 2MiB, 1024x1536

Anonymous Mon 22 May 2023 11:32:54 No.7389282 Report

Quoted By:

Anonymous

Anonymous Mon 22 May 2023 11:35:02 No.7389287 Report

Quoted By: >>7389508

Any tips to suppress extra human ears on animal eared characters? The LoRA /should/ suppress these but occasionally they peek out from behind her hair here and there and can be surprisingly difficult to inpaint out

Anonymous

Anonymous Mon 22 May 2023 11:35:26 No.7389290 Report

Quoted By:

>>7388920
i've had more success with hands using celeb/e-celeb loras and embeddings than any negative hand embeddings kek

Anonymous

Anonymous Mon 22 May 2023 11:36:19 No.7389292 Report

Quoted By:

>>7388769
There's numerous, but the ones I always go back to are:
AOM3 A2
AOM3 7th C
AOM3 Pastel 30 & 40
AOM3 Pastel 30 r34 SF1
AOM3 Nutmeg Ga 30 & 40
Based 64 & 65
Nabylon
Walnut Cream Blend Herb Mix

Anonymous

Anonymous Mon 22 May 2023 11:44:52 No.7389301 Report

Quoted By: >>7389305

>>7389149
what exactly does the padding slider do?

Anonymous

Anonymous Mon 22 May 2023 11:47:06 No.7389305 Report

Quoted By: >>7389310 >>7389362

>>7389301
It pads.
Its just influencing how many pixels the AI "sees". If you set it to a low number it pretty much has no clue whats around it and thus can make a face where you'd normally want a hand, because it doesn't know this is the end of an arm.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00038.png, 786KiB, 2560x3840

Anonymous Mon 22 May 2023 11:48:34 No.7389307 Report

Quoted By: >>7389316

feels nice to take a break from training

Anonymous

Anonymous Mon 22 May 2023 11:50:53 No.7389310 Report

Quoted By:

>>7389305
holy shit i finally understand, i've been using it with only a vague guess, thanks

Anonymous

Anonymous Mon 22 May 2023 11:50:59 No.7389311 Report

Quoted By: >>7389325

>>7389234
most of his loras are good

Anonymous

Anonymous Mon 22 May 2023 11:53:56 No.7389316 Report

Quoted By: >>7389392

>>7389307
catbox?

Anonymous

View Same Google ImgOps iqdb SauceNAO 66874-2625305847-1girl, solo foc (...).png, 2MiB, 1536x1536

Anonymous Mon 22 May 2023 11:54:50 No.7389319 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO image_2023-05-22_155647797.png, 1MiB, 1280x1920

Anonymous Mon 22 May 2023 11:56:59 No.7389324 Report

Quoted By: >>7389336 >>7389452

insert coin to play

Anonymous

Anonymous Mon 22 May 2023 11:57:30 No.7389325 Report

Quoted By:

>>7389311
no

Anonymous

Anonymous Mon 22 May 2023 12:03:18 No.7389336 Report

Quoted By: >>7389359

>>7389324
what are we gonna play

Anonymous

Anonymous Mon 22 May 2023 12:04:22 No.7389338 Report

Quoted By: >>7389355 >>7389424

>>7389109
i do it, doubt anyone cares though but its just convenient for me to have information and an example embeded in the LoRA file itself

Anonymous

View Same Google ImgOps iqdb SauceNAO spider.png, 561KiB, 832x1152

Anonymous Mon 22 May 2023 12:06:32 No.7389343 Report

Quoted By: >>7389413

Puffiest

Anonymous

Anonymous Mon 22 May 2023 12:09:13 No.7389347 Report

Quoted By: >>7389361

what was the prompting format to merge two prompts together, for example to get a leotard sailor collar? [leotard:sailor collar:0.4]? something like that?

Anonymous

Anonymous Mon 22 May 2023 12:11:54 No.7389355 Report

Quoted By: >>7389392

>>7389338
if you're the lazylora trainer anon I care

Anonymous

Anonymous Mon 22 May 2023 12:12:16 No.7389359 Report

Quoted By:

>>7389336
Battle City

Anonymous

Anonymous Mon 22 May 2023 12:13:25 No.7389361 Report

Quoted By:

>>7389347
Technically that's the prompt for alternating between sampling iterations [tag1,tag2:str] which has the effect of merging things but there's no reason to do it in your example because they're separate objects.

Anonymous

Anonymous Mon 22 May 2023 12:13:32 No.7389362 Report

Quoted By: >>7389371 >>7389409

>>7389305
Is there ANY reason it's set so low as default?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00231-3115600398-nsfw, scan, ((( (...).png, 2MiB, 1024x1536

Anonymous Mon 22 May 2023 12:14:46 No.7389366 Report

Quoted By: >>7389392

Anonymous

Anonymous Mon 22 May 2023 12:17:03 No.7389371 Report

Quoted By:

>>7389362
Dunno, but its not like high = always better.
If you set it too high the AI will not let you draw what you want, like drawing a new person in the background or making some localized change which seems nonsensical to AI. Drawing new faces on some dude in the crowd, idk. Pretty sure keeping it also impacts styles - if you set it higher the style will be nudged towards what the rest of the image looks like

Anonymous

View Same Google ImgOps iqdb SauceNAO 00040.png, 1MiB, 2560x3840

Anonymous Mon 22 May 2023 12:29:05 No.7389392 Report

Quoted By: >>7389398

>>7389316
https://files.catbox.moe/8odwjn.png
lora isn't released, though it shouldn't be too dissimilar to the current hades lora i have on gayshit. i used the pixelization extension for the effect
>>7389355
nta you replied to, but i am that trainer and yeah, it is convenient because i honestly forget probably like half of the loras i make lol
>>7389366
based pawpose

Anonymous

Anonymous Mon 22 May 2023 12:32:23 No.7389398 Report

Quoted By:

>>7389392
oh, hades. I knew the style's familiar and was curious what it was, thanks.

Anonymous

Anonymous Mon 22 May 2023 12:33:35 No.7389399 Report

Quoted By: >>7389491

Wtf is vae batch size? What do I set it to?

Anonymous

Anonymous Mon 22 May 2023 12:51:58 No.7389408 Report

Quoted By: >>7389418 >>7389427

Is there some kind of VRAM leak with this shit? I notice I got from being to upscale to 1.5k vertical to not even being able to upscale to 1k after enough generations. Is that vlad repo any better with this? I'm still using a1111

Anonymous

View Same Google ImgOps iqdb SauceNAO NewCanvas1.jpg, 1MiB, 5000x5000

Anonymous Mon 22 May 2023 12:53:57 No.7389409 Report

Quoted By:

>>7389362
you have a limited canvas size.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00122-548633520.png, 2MiB, 1024x1536

Anonymous Mon 22 May 2023 12:59:41 No.7389413 Report

Quoted By:

>>7389343
Is this a Lora? There's not even a tag for puffy on danbooru.

Anonymous

View Same Google ImgOps iqdb SauceNAO xyz_grid-0071-1787492555-solo, N (...).jpg, 401KiB, 4000x493

Anonymous Mon 22 May 2023 12:59:53 No.7389414 Report

Quoted By: >>7389417

...
there still seems to be some bizarre stair issue. I give up though, doesnt always happen and can be mitigated via weights probably.
How do people deal with such background overfit in the first place

Anonymous

Anonymous Mon 22 May 2023 13:01:25 No.7389417 Report

Quoted By: >>7389423

>>7389414
lower your lora dim/resize it or else ABG remover/paint over with some white or whatever and use inpaint around the borders to fix.

Anonymous

Anonymous Mon 22 May 2023 13:02:22 No.7389418 Report

Quoted By:

>>7389408
There can be leaks, but most likely you just rolled a higher VRAM gen this one time.

Anonymous

Anonymous Mon 22 May 2023 13:05:32 No.7389423 Report

Quoted By: >>7389425

>>7389417
Would have to nuke all BGs so not very realistic to clean it up.
I guess dims are the obvious solution, but dont most maintain 128/128?

Anonymous

Anonymous Mon 22 May 2023 13:05:52 No.7389424 Report

Quoted By:

>>7389338
i care

Anonymous

Anonymous Mon 22 May 2023 13:14:12 No.7389425 Report

Quoted By: >>7389468

>>7389423
most seem to resize dim to 64 or 32
I still don't really understand what alpha is/does but I train at 128dim/64alpha

Anonymous

Anonymous Mon 22 May 2023 13:15:04 No.7389427 Report

Quoted By:

>>7389408
You can maybe try adding
set PYTORCH_CUDA_ALLOC_CONF=garbage_collection_threshold:0.9,max_split_size_mb:512
to webui-user.bat
it's from the a1111 wiki under optimizations
>No performance impact and increases initial memory footprint a bit but reduces memory fragmentation in long runs
I feel like it helped a bit for me (8gb vramlet) but I didn't really do any testing

Anonymous

View Same Google ImgOps iqdb SauceNAO 00000-847495551.png, 2MiB, 1024x1536

Anonymous Mon 22 May 2023 13:20:49 No.7389434 Report

Quoted By: >>7389447

when you and your friend go for a morning beach run, you know how it goes right anon?

Anonymous

Anonymous Mon 22 May 2023 13:38:17 No.7389447 Report

Quoted By: >>7389532

>>7389434
Yeah, I collapse after 5 minutes and she laughs at how out of shape I am.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00007-1009628266.0.jpg, 2MiB, 2560x3840

Anonymous Mon 22 May 2023 13:43:41 No.7389451 Report

Quoted By: >>7389455

warrior (priest?) sluts

Anonymous

Anonymous Mon 22 May 2023 13:44:18 No.7389452 Report

Quoted By:

>>7389324
paizuri

Anonymous

Anonymous Mon 22 May 2023 13:46:47 No.7389455 Report

Quoted By: >>7389472

>>7389451
Looks like a priest to me
Did you inpaint hands and the staff?

Anonymous

Anonymous Mon 22 May 2023 13:48:19 No.7389456 Report

Quoted By:

>>7388486
but how the fuck are you supposed to actually use the prompt appending thing, there is no description of what that does

Anonymous

View Same Google ImgOps iqdb SauceNAO 00414-2200925174.png, 4MiB, 1536x2304

Anonymous Mon 22 May 2023 14:00:38 No.7389466 Report

Quoted By:

careful

Anonymous

Anonymous Mon 22 May 2023 14:05:43 No.7389468 Report

Quoted By: >>7389474 >>7389595

>>7389425
alpha is a multiplier of sorts.
resizing fucks up style so meh

Anonymous

View Same Google ImgOps iqdb SauceNAO 00059-3607343948.png, 1MiB, 1024x1536

Anonymous Mon 22 May 2023 14:10:56 No.7389472 Report

Quoted By:

>>7389455
just the hand with the staff. photoshopped then blended with inpaint

Anonymous

Anonymous Mon 22 May 2023 14:11:57 No.7389474 Report

Quoted By: >>7389484

>>7389468
>resizing fucks up style so meh
not really from what I've seen

Anonymous

Anonymous Mon 22 May 2023 14:12:05 No.7389475 Report

Quoted By:

>>7389036
>>7389038
>>7389056
https://files.catbox.moe/0dt4bx.png

Anonymous

Anonymous Mon 22 May 2023 14:15:39 No.7389484 Report

Quoted By:

>>7389474
look harder then

Anonymous

Anonymous Mon 22 May 2023 14:19:43 No.7389491 Report

Quoted By:

>>7389399
At least 8.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00096-2991778280.png, 3MiB, 1240x1240

Anonymous Mon 22 May 2023 14:25:21 No.7389495 Report

Quoted By:

Anonymous

Anonymous Mon 22 May 2023 14:33:43 No.7389501 Report

Quoted By:

>>7388940
i will try it out later, i've been using your stuff for a while and appreciate the work

Anonymous

Anonymous Mon 22 May 2023 14:44:40 No.7389508 Report

Quoted By: >>7389509

>>7389287
put ears in negatives and weight it up. hair over ears works too

Anonymous

Anonymous Mon 22 May 2023 14:46:53 No.7389509 Report

Quoted By: >>7389512 >>7389514

>>7389508
>put ears in negatives
I've done this but I wasn't sure if it actually had an effect since it seemed like the kind of thing that isn't usually tagged...
>hair over ears
That's a good idea. I've tried weighting up "bangs" or "animal ears" (usually not necessary since baked into the LoRA but I figured maybe it might reduce the occurrence of normal ears) I had a gen I did recently that was REALLY stubborn about inserting a little peek of a human ear behind the hair in one part I eventually just gave up and photoshopped out though. Maybe also depends on the model I'm using...

Anonymous

Anonymous Mon 22 May 2023 14:49:19 No.7389512 Report

Quoted By:

>>7389509
sidelocks will probably work better

Anonymous

Anonymous Mon 22 May 2023 14:50:01 No.7389514 Report

Quoted By: >>7389518

>>7389509
hair covering ears is good too. it's also really fucking easy to guided inpaint ears away just put some color in there done

Anonymous

Anonymous Mon 22 May 2023 14:53:25 No.7389518 Report

Quoted By: >>7389520

>>7389514
I haven't been able to use inpaint sketch in a while since it just instantly lags my browser to death any time I try it ever since some update (unless you're talking about something else)

Maybe I'll try in another browser though could be an issue with my Firefox (since I fucked with some cache related settings which cause issues occasionally on stuff like youtube)

Anonymous

Anonymous Mon 22 May 2023 14:54:28 No.7389520 Report

Quoted By: >>7389531

>>7389518
nah i meant take image to photoshop -> eyedropper the hair color -> brush over the ear -> put in inpaint -> done

Anonymous

Anonymous Mon 22 May 2023 14:59:30 No.7389531 Report

Quoted By:

>>7389520
Oh, yeah, I did that... and the ear fragment immediately came back. I think it might be an issue with the model or pose or something where it just really wanted to put one into that position no matter what

Anonymous

View Same Google ImgOps iqdb SauceNAO 00001-922752832.png, 1018KiB, 768x1152

Anonymous Mon 22 May 2023 15:00:14 No.7389532 Report

Quoted By:

>>7389447
then she kicks off her clothes and continues mocking you for being a fat fuck, bragging how she's so in shape that even her pussy muscles are stronger than you

she amuses herself by flexing it while talking in a high-pitched austrian accent, unaware that you just had a heart attack two minutes ago

Anonymous

Anonymous Mon 22 May 2023 15:11:31 No.7389544 Report

Quoted By: >>7389597

>>7386469
damn these lines are clean. You touch it up at all or is that strait genning?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00209-3201386027-nsfw, scan, ((t (...).png, 2MiB, 1024x1536

Anonymous Mon 22 May 2023 15:39:59 No.7389571 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO -__-.png, 427KiB, 512x512

Anonymous Mon 22 May 2023 15:48:03 No.7389577 Report

Quoted By: >>7389579 >>7389587

>mfw checking civishit for new checkpoints and seeing that 99% of those are for generating realistic 1girls with korean/chinese/asian face features

Anonymous

Anonymous Mon 22 May 2023 15:50:23 No.7389579 Report

Quoted By:

>>7389577
That's because it's what the real artist does. Everyone is a coomer.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00377-3475065715.png, 2MiB, 1024x1536

Anonymous Mon 22 May 2023 16:01:57 No.7389587 Report

Quoted By: >>7389619

>>7389577
it's because there is that stupid asian model instagram dataset out there, everyone used it for 3d finetunes now they can only generate korean sameface

Anonymous

View Same Google ImgOps iqdb SauceNAO catte.png, 997KiB, 768x1152

Anonymous Mon 22 May 2023 16:08:29 No.7389592 Report

Quoted By:

Anonymous

View Same Google ImgOps iqdb SauceNAO catte3.png, 1MiB, 768x1152

Anonymous Mon 22 May 2023 16:09:01 No.7389593 Report

Quoted By:

Anonymous

Anonymous Mon 22 May 2023 16:09:52 No.7389595 Report

Quoted By: >>7389734

>>7389468
>resizing fucks up style so meh
It doesn't unless you go overboard.

Anonymous

Anonymous Mon 22 May 2023 16:10:54 No.7389597 Report

Quoted By:

>>7389544
Just genning, I do not know how to inpaint or use img2img effectively yet IoI

Anonymous

View Same Google ImgOps iqdb SauceNAO 00156-2295646828.png, 2MiB, 1168x1440

Anonymous Mon 22 May 2023 16:20:46 No.7389600 Report

Quoted By: >>7389734

>captcha: NTRK0
kek

Anonymous

View Same Google ImgOps iqdb SauceNAO image_2023-05-22_122918004.png, 635KiB, 768x768

Anonymous Mon 22 May 2023 16:29:34 No.7389612 Report

Quoted By: >>7389619 >>7389967

Anonymous

View Same Google ImgOps iqdb SauceNAO devil.png, 3MiB, 1440x2160

Anonymous Mon 22 May 2023 16:30:38 No.7389614 Report

Quoted By: >>7389615 >>7389616 >>7389619 >>7389621 >>7389624 >>7389638 >>7389646 >>7389658

new thread:

>>7389611
>>7389611
>>7389611

Anonymous

Anonymous Mon 22 May 2023 16:31:44 No.7389615 Report

Quoted By:

>>7389614
huh

Anonymous

Anonymous Mon 22 May 2023 16:31:54 No.7389616 Report

Quoted By: >>7389661

>>7389614
>retards making threads when there's only 261 images
this place really is full of newfag niggers now

Anonymous

View Same Google ImgOps iqdb SauceNAO 00238-2415478238.png, 2MiB, 1024x1536

Anonymous Mon 22 May 2023 16:32:21 No.7389619 Report

Quoted By:

>>7389587
great giga benis

>>7389612
nice character

>>7389614
y tho?

Anonymous

Anonymous Mon 22 May 2023 16:32:54 No.7389621 Report

Quoted By: >>7389672

>>7389614
even if you made this early for "correcting" the OP, you already fucked it up. it didn't fix the catbox script link.

Anonymous

Anonymous Mon 22 May 2023 16:33:20 No.7389624 Report

Quoted By:

>>7389614
retard

Anonymous

Anonymous Mon 22 May 2023 16:45:36 No.7389638 Report

Quoted By:

>>7389614
retard

Anonymous

View Same Google ImgOps iqdb SauceNAO 1684324769307128-3144423609-nsfw (...).png, 1MiB, 768x1152

Anonymous Mon 22 May 2023 16:49:19 No.7389645 Report

Quoted By:

Anonymous

Anonymous Mon 22 May 2023 16:50:12 No.7389646 Report

Quoted By:

>>7389614
why?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00052-3328322986.png, 2MiB, 1280x1280

Anonymous Mon 22 May 2023 16:56:19 No.7389657 Report

Quoted By: >>7389668

Anonymous

Anonymous Mon 22 May 2023 16:58:07 No.7389658 Report

Quoted By:

>>7389614
fuck off

Anonymous

Anonymous Mon 22 May 2023 17:02:51 No.7389661 Report

Quoted By: >>7389663 >>7389672

>>7389616
>this place really is full of newfag n-words now

I don't see what the problem is.

Anonymous

Anonymous Mon 22 May 2023 17:03:40 No.7389663 Report

Quoted By:

>>7389661
kys baiting nigger

Anonymous

Anonymous Mon 22 May 2023 17:08:17 No.7389667 Report

Quoted By: >>7389675

I will now bake dadapt with constant

Anonymous

Anonymous Mon 22 May 2023 17:11:44 No.7389668 Report

Quoted By: >>7389678

>>7389657
Catbox for this?

Anonymous

View Same Google ImgOps iqdb SauceNAO 1684324769307148-2337529236-nsfw (...).png, 1MiB, 768x1152

Anonymous Mon 22 May 2023 17:12:13 No.7389671 Report

Quoted By: >>7389674

Anonymous

Anonymous Mon 22 May 2023 17:12:40 No.7389672 Report

Quoted By:

>>7389661
>page 7
>around 40 more images until the limit
with how slow this thread is sometimes it can take us half a day to get to the cap. There was literally no reason to make a new thread this early

>>7389621
good thing anon has plenty of time to delete and re-do it properly

Anonymous

Anonymous Mon 22 May 2023 17:13:14 No.7389674 Report

Quoted By: >>7389681

>>7389671
nice eyes, model/lora?

Anonymous

Anonymous Mon 22 May 2023 17:14:20 No.7389675 Report

Quoted By:

>>7389667
MRS OBAMA GET DOWN

Anonymous

View Same Google ImgOps iqdb SauceNAO 00070-2598529569.png, 1MiB, 1280x1280

Anonymous Mon 22 May 2023 17:15:46 No.7389678 Report

Quoted By:

>>7389668
https://files.catbox.moe/btur4a.png

Anonymous

View Same Google ImgOps iqdb SauceNAO 01233-1051619327.png, 2MiB, 1024x1536

Anonymous Mon 22 May 2023 17:17:06 No.7389680 Report

Quoted By: >>7389683

Anonymous

Anonymous Mon 22 May 2023 17:17:14 No.7389681 Report

Quoted By: >>7389685

>>7389674
is just style lora, good artist
https://files.catbox.moe/lv02ho.png

Anonymous

Anonymous Mon 22 May 2023 17:18:12 No.7389683 Report

Quoted By:

>>7389680
based

Anonymous

Anonymous Mon 22 May 2023 17:20:31 No.7389685 Report

Quoted By: >>7389692 >>7389698 >>7389848

>>7389681
thanks
>40010prototype_v4-000008(4fe004622f9f),
can't find this on gayshit/archive/civit, can you post pls?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00128-382827404-.png, 1MiB, 1024x1552

Anonymous Mon 22 May 2023 17:31:54 No.7389691 Report

Quoted By: >>7389693 >>7389695

anyone else rather fond of women?

Anonymous

Anonymous Mon 22 May 2023 17:32:01 No.7389692 Report

Quoted By: >>7389848

>>7389685
it was posted in the thread but i cant find it hold on

Anonymous

View Same Google ImgOps iqdb SauceNAO 1684324769307152-1971875099-nsfw (...).png, 1MiB, 768x1152

Anonymous Mon 22 May 2023 17:32:34 No.7389693 Report

Quoted By:

>>7389691
yeah...

Anonymous

View Same Google ImgOps iqdb SauceNAO 00000-2574501693.png, 1MiB, 896x1344

Anonymous Mon 22 May 2023 17:34:17 No.7389694 Report

Quoted By:

Anonymous

Anonymous Mon 22 May 2023 17:34:46 No.7389695 Report

Quoted By:

>>7389691
no i hate em

Anonymous

View Same Google ImgOps iqdb SauceNAO 1684324769307155-4011966611-nsfw (...).png, 1MiB, 768x1152

Anonymous Mon 22 May 2023 17:36:20 No.7389698 Report

Quoted By: >>7389848

>>7389685
https://files.catbox.moe/p55gys.safetensors

Anonymous

View Same Google ImgOps iqdb SauceNAO catbox_b435s0.png, 1MiB, 1536x1024

Anonymous Mon 22 May 2023 17:39:21 No.7389699 Report

Quoted By:

Anonymous

Anonymous Mon 22 May 2023 18:17:30 No.7389734 Report

Quoted By: >>7389754

>>7389600
cool frame

>>7389595
resizing is not a magical solution, you're trimming the lora and it doesnt just remove the "bad" aspects, there is a difference between how it'll look even if done properly.

Anonymous

Anonymous Mon 22 May 2023 18:34:49 No.7389754 Report

Quoted By: >>7389807

>>7389734
>there is a difference between how it'll look even if done properly.
I disagree.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00094-4281068785.png, 2MiB, 1280x1280

Anonymous Mon 22 May 2023 18:51:06 No.7389786 Report

Quoted By: >>7389816 >>7389833

Anonymous

Anonymous Mon 22 May 2023 19:06:24 No.7389807 Report

Quoted By:

>>7389754
invest in glasses.

Anonymous

Anonymous Mon 22 May 2023 19:08:20 No.7389816 Report

Quoted By:

>>7389786
I think you succeeded this time. It finally looks far enough from lycoreco

Anonymous

View Same Google ImgOps iqdb SauceNAO 00285-111663527.png, 2MiB, 1536x1024

Anonymous Mon 22 May 2023 19:09:12 No.7389819 Report

Quoted By: >>7389832

Anonymous

View Same Google ImgOps iqdb SauceNAO 00737-3567296233.png, 393KiB, 1688x1120

Anonymous Mon 22 May 2023 19:11:20 No.7389824 Report

Quoted By:

>accidentally leave inpaint on fill
the cockhead

also why is generating a foreskin dick so goddamn difficult

Anonymous

View Same Google ImgOps iqdb SauceNAO 1684324769307215-208478032-nsfw, (...).png, 2MiB, 1024x1536

Anonymous Mon 22 May 2023 19:15:21 No.7389831 Report

Quoted By:

Anonymous

Anonymous Mon 22 May 2023 19:16:14 No.7389832 Report

Quoted By:

>>7389819
Based

Anonymous

Anonymous Mon 22 May 2023 19:16:58 No.7389833 Report

Quoted By:

>>7389786
>cute genshins with big breasts
You're doing god's work.

Anonymous

View Same Google ImgOps iqdb SauceNAO 1684324769307216-2020739441-nsfw (...).png, 2MiB, 1024x1536

Anonymous Mon 22 May 2023 19:21:10 No.7389839 Report

Quoted By:

controlnet is fun

Anonymous

View Same Google ImgOps iqdb SauceNAO catbox_cqqy3p.png, 2MiB, 1024x1536

Anonymous Mon 22 May 2023 19:21:44 No.7389840 Report

Quoted By: >>7389867

Anonymous

View Same Google ImgOps iqdb SauceNAO 00023-3945322385.png, 2MiB, 1024x1536

Anonymous Mon 22 May 2023 19:25:26 No.7389848 Report

Quoted By:

>>7389692
>>7389685
>>7389698
Not sure why it's not on gayshit but here's the link I initially posted
https://mega.nz/folder/SAIgTCCC#bvDU4A3iEkbns4TDnx7-Qg

artist-40010prototype (Shimanto Shisakugata)

Anonymous

Anonymous Mon 22 May 2023 19:40:11 No.7389866 Report

Quoted By:

owari da

Anonymous

View Same Google ImgOps iqdb SauceNAO 01762-2033715769.png, 2MiB, 1536x1024

Anonymous Mon 22 May 2023 19:40:54 No.7389867 Report

Quoted By: >>7389869 >>7389871

>>7389840
Speaking of sex machines, I like the idea of a mecha cockpit with a sex machine built into, like it's necessary to power or control the mecha for whatever chuuni reason.

Anonymous

Anonymous Mon 22 May 2023 19:42:14 No.7389869 Report

Quoted By:

>>7389867
it's pretty good at that without assistance, cockpit is a strong aesthetic

Anonymous

View Same Google ImgOps iqdb SauceNAO 01778-2700887152.png, 2MiB, 1536x1024

Anonymous Mon 22 May 2023 19:42:41 No.7389871 Report

Quoted By: >>7389876 >>7389886

>>7389867
The problem is, I'd prefer to have the girl hold onto joysticks built into the arms of the pilot seat, but SD doesn't understand that concept. If I try to prompt "both hands on joysticks", it gives me a game controller.

Anonymous

Anonymous Mon 22 May 2023 19:45:19 No.7389876 Report

Quoted By:

>>7389871
>If I try to prompt "both hands on joysticks", it gives me a game controller.
based AI

Anonymous

Anonymous Mon 22 May 2023 19:59:01 No.7389886 Report

Quoted By:

>>7389871
try some dva lora on low strength just for the concept thats prob baked into it?

Anonymous

View Same Google ImgOps iqdb SauceNAO now which cookie☆ character do y (...).jpg, 433KiB, 1280x1536

Anonymous Mon 22 May 2023 20:03:33 No.7389893 Report

Quoted By: >>7389895

I'm starting to get more enjoyment out of contributing to the webui and other repos to help other anons generate images than generating images myself, whoops.

Anonymous

Anonymous Mon 22 May 2023 20:04:27 No.7389895 Report

Quoted By: >>7389955

>>7389893
Im a long time touhoufag and I still don't understand what the fuck cookie is

Anonymous

View Same Google ImgOps iqdb SauceNAO 1684324769307231-4182645979-nsfw (...).png, 1MiB, 768x1152

Anonymous Mon 22 May 2023 20:05:54 No.7389898 Report

Quoted By: >>7389908

Anonymous

View Same Google ImgOps iqdb SauceNAO 1684324769307235-4111673792-nsfw (...).png, 1MiB, 768x1152

Anonymous Mon 22 May 2023 20:07:41 No.7389900 Report

Quoted By:

Anonymous

Anonymous Mon 22 May 2023 20:12:21 No.7389908 Report

Quoted By: >>7389931

>>7389898
Nice. Catbox, please?

Anonymous

Anonymous Mon 22 May 2023 20:14:01 No.7389912 Report

Quoted By:

What is the difference between Locon and Locon_Lycoris

Anonymous

View Same Google ImgOps iqdb SauceNAO standby.png, 1MiB, 768x1152

Anonymous Mon 22 May 2023 20:20:43 No.7389931 Report

Quoted By: >>7389935

>>7389908
arm prompts are gacha af

https://files.catbox.moe/hwj156.png

Anonymous

Anonymous Mon 22 May 2023 20:23:22 No.7389935 Report

Quoted By:

>>7389931
Cheers, mate

Anonymous

View Same Google ImgOps iqdb SauceNAO 00008-2239133886-1girl, (((from (...).png, 2MiB, 1560x2184

Anonymous Mon 22 May 2023 20:28:23 No.7389938 Report

Quoted By: >>7389939 >>7389941

>this actually works
that one kneepitussy bro from before boutta bust some nuts

Anonymous

Anonymous Mon 22 May 2023 20:29:12 No.7389939 Report

Quoted By: >>7389958

>>7389938
make the pits lewder. We have the technology.

Anonymous

Anonymous Mon 22 May 2023 20:32:05 No.7389941 Report

Quoted By:

>>7389938
SMELLY SWEATY HAIRY KNEEPUSSY

Anonymous

Anonymous Mon 22 May 2023 20:39:35 No.7389955 Report

Quoted By:

>>7389895
Same

Anonymous

Anonymous Mon 22 May 2023 20:41:02 No.7389958 Report

Quoted By: >>7389962 >>7389963

>>7389939
Isn't this what you wanted? Why are you afraid?
https://files.catbox.moe/fr1d15.png

Anonymous

Anonymous Mon 22 May 2023 20:42:43 No.7389959 Report

Quoted By:

how do you change preprocessor parameters on controlnet?

Anonymous

View Same Google ImgOps iqdb SauceNAO ENHANCE_1.jpg, 29KiB, 479x433

Anonymous Mon 22 May 2023 20:43:34 No.7389962 Report

Quoted By:

>>7389958

Anonymous

Anonymous Mon 22 May 2023 20:44:41 No.7389963 Report

Quoted By:

>>7389958
too lewd...

Anonymous

View Same Google ImgOps iqdb SauceNAO 1684324769307258-484189308-cowbo (...).png, 1MiB, 768x1152

Anonymous Mon 22 May 2023 20:46:01 No.7389965 Report

Quoted By:

Anonymous

Anonymous Mon 22 May 2023 20:49:27 No.7389967 Report

Quoted By:

>>7389612
gigabased

Anonymous

View Same Google ImgOps iqdb SauceNAO 00047-1318091021.png, 1MiB, 1024x1280

Anonymous Mon 22 May 2023 20:55:48 No.7389973 Report

Quoted By: >>7389974 >>7390036 >>7390045

Anonymous

Anonymous Mon 22 May 2023 20:56:26 No.7389974 Report

Quoted By: >>7389977

>>7389973
ga-

Anonymous

View Same Google ImgOps iqdb SauceNAO yellow.png, 1MiB, 768x1152

Anonymous Mon 22 May 2023 21:00:31 No.7389976 Report

Quoted By:

Anonymous

Anonymous Mon 22 May 2023 21:00:32 No.7389977 Report

Quoted By: >>7389979 >>7390037

>>7389974
https://files.catbox.moe/x6c0zm.png
Not really what was prompted though
I wonder if it's possible to prompt the "below spread legs" pov

Anonymous

View Same Google ImgOps iqdb SauceNAO 00310-2793378240.png, 2MiB, 1024x1536

Anonymous Mon 22 May 2023 21:00:42 No.7389978 Report

Quoted By:

Anonymous

Anonymous Mon 22 May 2023 21:01:47 No.7389979 Report

Quoted By: >>7390000 >>7390018

>>7389977
>I wonder if it's possible to prompt the "below spread legs" pov
Doubt it, but I've gotten some umpromtables by sheer RNG.

Anonymous

View Same Google ImgOps iqdb SauceNAO 07519-2485048901-nsfw, masterpie (...).png, 1005KiB, 768x1152

Anonymous Mon 22 May 2023 21:18:42 No.7389988 Report

Quoted By:

>>7389064
https://files.catbox.moe/5nutgt.png

Anonymous

View Same Google ImgOps iqdb SauceNAO 00012-1795716818.png, 379KiB, 512x640

Anonymous Mon 22 May 2023 21:36:05 No.7390000 Report

Quoted By: >>7390010 >>7390018 >>7390045

>>7389979
it actually seems to be possible but the anatomy gets completely fucked for most of the gens

Anonymous

Anonymous Mon 22 May 2023 21:47:45 No.7390006 Report

Quoted By:

>>7387757
Wait so you draw the pickle AND also describe what it's supposed to be or how does that work?

Anonymous

View Same Google ImgOps iqdb SauceNAO 00397-1998730838.png, 2MiB, 1024x1536

Anonymous Mon 22 May 2023 21:48:08 No.7390007 Report

Quoted By:

Anonymous

Anonymous Mon 22 May 2023 21:52:36 No.7390010 Report

Quoted By: >>7390027

>>7390000
What prompt prompted this pic?

Anonymous

View Same Google ImgOps iqdb SauceNAO 06922-3465204287-1girl, from_bel (...).png, 427KiB, 512x512

Anonymous Mon 22 May 2023 21:58:30 No.7390018 Report

Quoted By: >>7390033 >>7390036 >>7390048

>>7390000
>>7389979

It's not hard at all.
Just use "from_below" in combination with "1girl".

"feet_out_of_frame" probably also helps.

Anonymous

Anonymous Mon 22 May 2023 22:05:03 No.7390027 Report

Quoted By: >>7390030

>>7390010
1girl, hakurei reimu, (panties focus, from below,standing:1.4)
Negative prompt: (worst quality, low quality:1.2), looking at viewer

Upskirt seems to kinda help as well

Anonymous

Anonymous Mon 22 May 2023 22:08:33 No.7390030 Report

Quoted By: >>7390032

>>7390027
What do the numbers mean? 1.4 1.2 etc?

Anonymous

Anonymous Mon 22 May 2023 22:09:43 No.7390032 Report

Quoted By: >>7390035

>>7390030
https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#attentionemphasis

Anonymous

View Same Google ImgOps iqdb SauceNAO 06935-1363170528-1girl, (from be (...).png, 353KiB, 512x512

Anonymous Mon 22 May 2023 22:10:06 No.7390033 Report

Quoted By: >>7390036 >>7390048

>>7390018
Actually judging from a couple of gens, it's better without the feet prompt.

Anonymous

Anonymous Mon 22 May 2023 22:13:17 No.7390035 Report

Quoted By:

>>7390032
thank

Anonymous

Anonymous Mon 22 May 2023 22:14:04 No.7390036 Report

Quoted By: >>7390045

>>7390018
>>7390033
Not really >>7389973 tho. Try with close up.

Anonymous

View Same Google ImgOps iqdb SauceNAO 06939-448028746-1girl, (from bel (...).png, 347KiB, 512x512

Anonymous Mon 22 May 2023 22:15:15 No.7390037 Report

Quoted By: >>7390039 >>7390048

>>7389977
So yeah, use (from below)

It works better with fewer other prompts behind it. I find that signifying skirt or panties color makes it more likely that you get more of a candid upskirt picture where you see her face. A simpler prompt gives the pure upskirt pics.

Anonymous

View Same Google ImgOps iqdb SauceNAO 06938-448028745-1girl, (from bel (...).png, 402KiB, 512x512

Anonymous Mon 22 May 2023 22:16:35 No.7390039 Report

Quoted By:

>>7390037

Anonymous

Anonymous Mon 22 May 2023 22:18:56 No.7390045 Report

Quoted By:

>>7390036
I was trying to recreate >>7390000
not >>7389973

Anonymous

View Same Google ImgOps iqdb SauceNAO grid-0005.png, 1MiB, 1024x1280

Anonymous Mon 22 May 2023 22:20:40 No.7390048 Report

Quoted By: >>7390050 >>7390051

>>7390037
>>7390033
>>7390018
Yeah those are achievable but I wanted a from below angle with legs straight , right under panties. Something like on 4th pic but not completely fucked up kek

Anonymous

View Same Google ImgOps iqdb SauceNAO 06946-2834039969-1girl, (from be (...).png, 317KiB, 512x512

Anonymous Mon 22 May 2023 22:24:16 No.7390050 Report

Quoted By: >>7390051

>>7390048
Yeah you get some funny results.

Anonymous

View Same Google ImgOps iqdb SauceNAO 00069-2724738986-1girl, mature f (...).png, 2MiB, 1280x1792

Anonymous Mon 22 May 2023 22:25:05 No.7390051 Report

Quoted By:

>>7390048
I have still yet to recreate this accidental cunnilingus POV and there's no images to make a lora out of.
It's over...
>>7390050
POV: Ant

Anonymous

Anonymous Mon 22 May 2023 22:33:50 No.7390067 Report

Quoted By:

migrate
>>7389611

Anonymous

Anonymous Tue 23 May 2023 02:02:14 No.7390664 Report

Quoted By: >>7391029

>>7388753
learn to draw

Anonymous

Anonymous Tue 23 May 2023 07:46:47 No.7391029 Report

Quoted By:

>>7390664
Don't train in baboon faces.

Anonymous

Anonymous Tue 23 May 2023 12:00:03 No.7391163 Report

Quoted By:

>>7386170
UOHHHHHHHHH EROTIC!!!!!!!

Anonymous

Anonymous Wed 24 May 2023 05:01:56 No.7392268 Report

Quoted By:

>>7386193
I still cannot fathom how Japan continues to exist like this. We should drop leaflets of uncensored genitals and watch them lose their minds.

Anonymous

Anonymous Wed 24 May 2023 05:06:41 No.7392271 Report

Quoted By:

>>7386320
Hey that's big in Japan.

Anonymous

Anonymous Wed 24 May 2023 05:19:12 No.7392288 Report

Quoted By:

>>7386583
This is it. This is what I shall cum to tonight.

Anonymous

Anonymous Thu 25 May 2023 06:30:52 No.7394021 Report

Quoted By: >>7394309

This has got to be the biggest surviving thread in /h/ in the longest time.

Anonymous

Anonymous Thu 25 May 2023 12:55:28 No.7394309 Report

Quoted By:

>>7394021
Oops, didn't mean to say that. I mean, these Hentai Diffusion threads are really packed to the brim with stuff, content and posts.

Anonymous

Anonymous Fri 26 May 2023 10:44:39 No.7395810 Report

Quoted By:

Not a fan

https://www.twitch.tv/otzdarva

https://www.youtube.com/c/notOtzdarva

https://www.youtube.com/channel/UCDO__Z-tZD4bfG9m4AkK_uA

https://www.youtube.com/user/TrU3Ta1ent

https://www.twitch.tv/tru3ta1ent

https://www.youtube.com/channel/UCHklDPNc9WZFUY0jaZNdcFQ

Anonymous

Anonymous Fri 26 May 2023 16:55:07 No.7396137 Report

Quoted By:

So, I'm an artist thinking of getting into this stuff, fuck it, might as well. I'm getting that there's no more free options to run this? I tried paperspace but it prompts me to join pro to run a machine, and folks are saying google too is at capacity?

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 #325