[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / s / t / u / v / vg / vm / vmg / vr / vrpg / vst / w / wg] [i / ic] [r9k / s4s / vip] [cm / hm / lgbt / y] [3 / aco / adv / an / bant / biz / cgl / ck / co / diy / fa / fit / gd / hc / his / int / jp / lit / mlp / mu / n / news / out / po / pol / pw / qst / sci / soc / sp / tg / toy / trv / tv / vp / vt / wsg / wsr / x / xs] [Settings] [Search] [Mobile] [Home]
Board
Settings Mobile Home
/g/ - Technology


Thread archived.
You cannot reply anymore.


[Advertise on 4chan]


File: ADT.jpg (3.18 MB, 3854x2336)
3.18 MB JPG
Previous: >>109633770

>UIs
ComfyUI: https://github.com/comfyanonymous/ComfyUI
SwarmUI: https://github.com/mcmonkeyprojects/SwarmUI
KritaAI: https://github.com/Acly/krita-ai-diffusion
Forge Classic/Neo: https://github.com/Haoming02/sd-webui-forge-classic

>Image Models
IL: https://rentry.org/illustrious_checkpoints
Anima: https://huggingface.co/circlestone-labs/Anima
Krea: https://huggingface.co/krea/Krea-2-Turbo
NovelAI (paid): https://novelai.net

>Video Models
H3: https://huggingface.co/Comfy-Org/MiniMax-H3
Turbo lora: https://huggingface.co/lightx2v/Minimax-h3-Turbo/blob/main/minimax_h3_fl2v_turbo_4step_v1.1_768p_comfyui_bf16.safetensors

>How to generate anime:
image : https://rentry.org/comfyui_guide_1girl
video: https://rentry.org/wan22ldgguide (No H3 guide yet)

>Style Explorers
Anima: https://anima.mooshieblob.com/ | https://animadex.net/
Illustrious: https://tagexplorer.github.io

>Checkpoints, LoRAs, Upscalers, & Workflows
https://civitai.com
https://civitaiarchive.com
https://tensor.art
https://openmodeldb.info
https://www.seaart.ai
https://www.liblib.art/

>Resources:
Share Metadata: https://litterbox.catbox.moe
Img2Prompt: https://huggingface.co/spaces/fancyfeast/joy-caption-beta-one
Online metadata viewer SD/NovelAI: https://spell.novelai.dev
Catbox/Metadata Userscript: https://gist.github.com/catboxanon/ca46eb79ce55e3216aecab49d5c7a3fb
Output cleanup: https://rentry.org/RemovingDiffusionGunk , https://www.mediafire.com/file/vipr23exc5htmnt (batch processing python script)
Inpainting Guide from an Anon: https://files.catbox.moe/fbzsxb.jpg >>106520607

>/adt/ logo:
https://files.catbox.moe/pakfgt.png

>Beware inorganic activity
https://rentry.org/animanon
>>
>>109650172
>Too early and has schizo rentry.
C-
see me after class
>>
Post more BA moetron abomination made in Krea2 please
>>
>>109650192
What's "BA moetron abomination"?
>>
File: 1787085263302.png (835 KB, 1024x1024)
835 KB PNG
>>109650229
>>
>>
File: file.png (1.58 MB, 888x1184)
1.58 MB PNG
>>109650238
I don't know BA characters and Krea 2 Turbo seems to have very spotty knowledge of this girl. I did this using Image Editing mode (which is pretty janky/limited still)
>>
>>109650238
>>109650305
not exactly a moetron but I get what you mean
>>
>>109650191
yeah, it does have your rentry
what about it?
>>
>>109650992
you are like the pirate software of diffusion generals
>>
File: good morning.jpg (357 KB, 1024x1024)
357 KB JPG
>>
>>109651142
stop talking to your reflection, Julien
>>
File: Krea2_turbo_01851_.jpg (1.79 MB, 2368x1776)
1.79 MB JPG
>>109651142
I hope you realize that I don't really post here until I learned of the state of things
If I was OP it would be a single image also I rarely bake so please stop.
>>
Sota bros how are we feeling? Feels good to have a 32ch anime model at the tip of my fingers
>>
>>109650238
Wow this is the power of Krea? Maybe in 5 years things will be solved?
>>
does anyone have a link to the chroma krea anima merge?
>>
>>109651305
https://huggingface.co/lodestones/Kroma
>>
>>109651185
Yeah I know, but for the sake of keeping things orderly, let's leave Ani here and make this her shrine.
>>
File: output_patch.mp4 (3.74 MB, 1920x1056)
3.74 MB
3.74 MB MP4
>>109651840
Would be great if he did, or use /sdg/. I don't have any horse in this general at all. I don't understand the start shit and cry victim mentality when people push back that his camp seem to have it's really pathetic. They don't even post gens anymore...I don't understand it you're still supposed to enjoy the hobby regardless of what happens.
>>
>>109651840
>>109651938
can you please leave your circular schizo self replies to /ldg/ from now on?
>>
>>109651963
You don't contribute to any thread
>>
>>109650192
The first two I posted a few threads ago were by luck, the rest needed some poking to avoid the default yuuka/aris outfit, or saiba ears which gets old after you see it once.

"military clothes" will trigger Iroha though not on her own name. Plus "white hair" to get Plana.
>>
I don't have anything against Ani or Catjak personally. The reason I post here specifically is to avoid running into them over in /ldg/. The fact that he needs to come here and avatarfag, when he could just talk to Ani directly on Discord, really kills my motivation to keep posting in this thread, and I think a lot of other anons feel the same way. I don't want to keep feeding this schizo loop. Last thread was a total shitpost fest and I'm not here to be a foot soldier for Ani or Catjak, I don't care about that. That's not why I came here, and it's not why I picked up this hobby, I like anime and I like anime girls, I don't want to end up stuck in some faction.

With that said, I'm out of this thread.
>>
File: Tankmage.jpg (266 KB, 1024x1536)
266 KB JPG
>>
>109652574
so after crying and splitbaking because "the op doesn't have model links" "the op doesn't have nai" this is your new concern trolling huh
notice how you're pretending like "avatarfagging" is the reason for your melty but you said nothing whenever ani avatarfagged or samefagged lol
thank you for the blog post, it was very organic and very stealthy
>>
>>109652722
>nothing whenever ani avatarfagged or samefagged lol
when did that happen?
>>
File: misc_20_55.jpg (637 KB, 1792x1792)
637 KB JPG
>>
File: file.png (213 KB, 890x896)
213 KB PNG
>>109652871
>avatarfagged
https://desuarchive.org/g/search/filename/anistudio%2A/
>literal hundreds of images sent with the same original character
gee anon i wonder
>samefagged
pic related just one example among many of your attempts to inorganically influence public opinion of yourself on our general
>>
File: Tier Harribel _2.jpg (1.85 MB, 2496x3648)
1.85 MB JPG
>>
Ok
I'm done
it's all one dude btw, he dropped the ball once on something really obvious and I'm autistic so that stuck with me

I hope you heal man x I really do.
>>
File: frierendoomtext.png (2.25 MB, 1024x1536)
2.25 MB PNG
>>
>>109652916
Why are you posting screenshots of your posts?
>>
File: ComfyUI_temp_urdqe_00010_.png (3.29 MB, 1024x1536)
3.29 MB PNG
Just testing to see if Krea can do sharply mixed styles (it can)
>>
>>109653483
>when background intern dabs on the animation team
>>
>>109653066
lmao

>>109653483
nice choice of artist, but she looks pasted into the image. maybe try making her interact with the environment more.

>>109653547
hot
>>
File: Bonk (For the Emperor).jpg (537 KB, 1536x1024)
537 KB JPG
>>
4.5 full gens at large and wallpaper resolution >>>>>>>>>32ch v5 slop
>>
File: misc_20_57.jpg (664 KB, 2176x1472)
664 KB JPG
>>
>>109654011
Neat, which krita fork are you using woth beter gui?
>>
after using anima base at 40 steps again I can't go back to turboslop
>>
File: ComfyUI_10789_.png (2.69 MB, 1248x1824)
2.69 MB PNG
>>109653687
Oh it was just a test using two polar opposite styles, if I were to do a concept like that seriously I would add more complex lighting to the subject, though in my opinion there's nothing wrong with a 2D-on-3D look if done tastefully
>>
File: ComfyUI_temp_dekyx_00052_.png (2.62 MB, 1024x1536)
2.62 MB PNG
>>109654275
Some styles do look unironically better with the turbo lora though
>>
>>109654525
what style?
>>
>>109654525
styles feel undercooked on turbo for me
prompt adherence and backgrounds take a hit too
>>
>>
File: salter vs honami.png (1.84 MB, 1216x832)
1.84 MB PNG
>>
>>
>>109655144
You think chess is bad, try tic-tac-toe to really throw her for a loop
>>
>>109655144
saber has no chance here
even with 1 finger advantage
>>
thanks for bakering
>>
File: mad bitch.png (1.47 MB, 1216x832)
1.47 MB PNG
>>
File: major.jpg (392 KB, 1728x1152)
392 KB JPG
>>
File: Krea2_turbo_00129_.png (1.83 MB, 1448x1448)
1.83 MB PNG
Anything interesting for z-image turbo? I played with it a bit then forgot about it. Krea2 is fun but annoying how the model fights you at times.
>>
>>109655481
It's over for Z Image I had fun testing it's failed Z Anime and it's ahit the bed and it's SaaS
>>
>>109655314
Neat, local or sota?
>>
>>109651568
better than anima?
>>
>>109655557
It's from Lodestone, you already know what to expect (failbake)
>>
>>109655573
Does >>109653547 look like a failbake to you?
>>
>>109655577
looks like a fried mess, yes.
>>
>>109655590
post your gen, you won't
>>
>>109655577
Are you seriously asking that in this thread? Newfag or just retarded? It depends on the prompt, how much the author actually put into it, and how the model responded to that. An image alone, no prompt, tells you nothing. Draw whatever conclusion you want, it's worthless.
Otherwise you're no different from the shitmergers comparing output images between merges without ever posting the prompt. Useless comparisons from useless people. What matters is the balance between adherence and creation, how tightly it follows the prompt versus how much it builds beyond it, without going full hallucination mode, stupid faggot, ypu are a worthless slopper.
>>
>>109655577
I actually don't know what caused it to go sketchomodo like that but it was too cool of a pose to mess with
>1 finger = boolit
>all five fingers = bazooka
Must've walked in on her bathing
>>
>>109655622
benchod
>>
File: poor weight distribution.jpg (318 KB, 1344x1344)
318 KB JPG
>>
>>109655592
I wasn't going to bother but I said fuck it and put in pure bullshit for an NL prompt on v5 and it actually worked
>>
File: 1756609248081877.png (1.5 MB, 1216x832)
1.5 MB PNG
>>109655651
>>
File: ComfyUI_temp_zehfv_00094_.png (2.19 MB, 1024x1536)
2.19 MB PNG
>>109654752
https://civitai.red/models/2877827/magic-knight-rayearth-style-clamp-anima

>>109654882
Both of those are true but it's the sort of simplification from the undercooking that works in certain styles' favour, for example if you want cleaner, less sketchy patterns on things
>>
>>109655651
is she ok?
>>
>>109655745
idk how you can say turbo makes better styles when you just cheat with a style lora anyways
>>
>>109655752
worse has happened
she will manage
>>
>>109655773
>cheat with a style lora
What?
>>
>>109655785
loras are cope. they are destructive to the model and end up being stiff representations of the dataset
>>
>>109655813
yeah everyone should just be using raw models instead. Great models like WAI or plantmilk or rdbt!
>>
>>109655651
Rika if she got Satoko's growth spurt to the chest only
>>
>>109655678
>six fingers
I thought this was supposed to be the ultimate sota anime model
>>
thoughts on novelai v5
https://www.bilibili.com/video/BV1i4826KEyg
>>
>>109655813
I genuinely can't tell if you're being serious or not. If you're concerned about "cope" then why have you not picked up a pencil?
>>
>>109655874
Illegible moonrunes
Grok, translate for my simple monkeybrain
>>
>>
>>109655883
listen man. in a perfect world we would just have a base model or fine-tune that actually slaps and does a wider variety of artists than just danbooru. Anima just isn't going to get there with how broken training is for large fine tunes that would hit that spot. krea2 can train well on anime and if you are going to use loras anyways, use a smarter model with better quality and anatomy
>>
>>109656147
>those eras
kek, accurate
>>
>>109655651
Oh.. ;ɛ
>>
>>109655678
NAI ambassador, quick question. Close up amd upper body shots come out great. But anything from further away, like full body, looks bad and kind of broken at 1216x832.
>>
>>109656299
What is the question, fagguete?
>>
>>109653483
Is that the background from the artist who I think it is?
>>109653692
Cute Tako.
>>
>>109655651
the head should weigh more too
>>
>open new /adt/ thread
>search 'tho'
>0 results
Why, tho?
>>
>>109655290
/_ _ _/ _ _ _ _ >:^(
>>
File: 1780653269595959.png (1.25 MB, 1216x832)
1.25 MB PNG
>>
File: 260826071547_output-0001.png (3.22 MB, 1096x1914)
3.22 MB PNG
>>
File: Earlier than predicted.jpg (2.63 MB, 832x3648)
2.63 MB JPG
>>109656400
Mr.Tho~
>>109656322
Thanks!
>>
>>109656230
You didn't answer my question. If you're concerned about "cope" why are you using machine hallucinations to guess at your vision instead of drawing something exactly the way you want?

>>109656322
It's Beksinski yes
>>
File: misc_20_56.jpg (705 KB, 2176x1472)
705 KB JPG
>>109654192
My own. And I recommend making your own fork.

Use the original for some time, just to understand what you like/dislike about it. Then ask any coding agent to make the fixes you want. Or make the fixes yourself, it's not the best codebase, but perfectly workable.
>>
>>109656521
kek
i like your images with her, she's hot
>>
File: is coffe good for you?.jpg (395 KB, 1152x1728)
395 KB JPG
>>
File: Nagato-nee.jpg (568 KB, 1856x1280)
568 KB JPG
>>
>>109656691
>People will just see this lovely bote and say wife
>>
>>109656705
*door slams open*
>Commander! What are you doing? Why did you take Nagato-nee's clothes off? What is that between your legs? Why are you spreading Nagato-nee's legs? What's the meaning of it? Are you playing a game? Is it fun? Can I join? And, um, why did that thing between your legs just get even bigger when I said that...? Tell me, I want to know!
>>
>>109655813
Shhh, please don’t educate the tourists, that’s supposed to be a secret.
>>
File: Double Retrofit.jpg (685 KB, 1467x2144)
685 KB JPG
>>109656691
>>
>>109656869
Well, I see that the Sakura Empire have equipped their flagship with ballistics too
>>
File: Supportsu.jpg (759 KB, 1467x2144)
759 KB JPG
>>109656884
lol
>>
>v5 was trained in-house on 268,000 B200 GPU hours
Joking aside, seeing the images from NAIv5, if I were an illustrator I would want to kill myself. My profession no longer has a use case, only as a dataset maker bitch and working for NAI as a dataset contributor.
>>
>>109657059
268,000 GPU hours is compute time across many GPUs running in parallel, not one GPU running continuously — so the real answer depends on how many GPUs they used at once.
If it were a single GPU running nonstop: 268,000 ÷ 24 ≈ 11,167 days ≈ about 30.6 years.
More realistically, training runs like this use hundreds or thousands of GPUs simultaneously. For example:
With 500 GPUs in parallel: 268,000 ÷ 500 ÷ 24 ≈ 22 days
With 1,000 GPUs in parallel: ≈ 11 days
With 2,000 GPUs in parallel: ≈ 5.6 days
>>
>>109657059
slop wouldn't exist without artists
you and i can't exist without parents
no bully
>>
>>109656582
He's not trying to answer your question, he's trying to shoehorn in the "anima is a broken untrainable dead-end model" claim. Like he does every day, in every AI imagegen thread, for the past several months.
>>
>>109657059
Everything I see from NAI looks like saturated ass. Like pure AI slop from SD 1.5
>>
File deleted.
>>109657232
what? sd1.5 had the exact opposite problem of being extremely desaturated. Hell that was a problem that persisted through SDXL outside of extremely overfried 20CFG output.
>>
rip anima
https://civitai.red/models/2881657/wulver-krea-2-v01-alpha
>>
>>109657245
>everything loaded into character slot
>verbatim prompting
N E A T
>>
>>109657369
Based furries unlike noobcord troons, they know how to work organized and love in unison. Animetroons condemned to fail because egos and 1girl wars.
>>
>>109657073
Did you see it coming how fast did the "AI slop" concept stop being slop, and how we raised the bar? Everyone swore this stuff could never touch an artist's composition, expression, or texture, all that. NAI walked in and planted the flag in 2026.
>>
>>109657549
My fingers are better inpainted than yours ;ɛ
>>
File: Chunguslith Miniigan.jpg (388 KB, 1024x1536)
388 KB JPG
>>
>>
>>109657592
the moment publicly accessible slopping became a thing we pointed out it was only a matter of time before it got scary
and most of us figured 5ish years
we're not prophets but...
>>
>>109657521
it's clear text encoders matter a lot
I still don't get why anima only has a 0.6b one
>>
>>109657703
Tourist here, how do I make HQ sluts like this one? I don't know shit about doing pics because whenever writing a prompt on Perchance, it fucks up the anatomy.
>>
>>109657059
this tech exists thanks to based illustrators
I wouldn't mind seeing western troon artists get replaced though since they don't produce much of value anyway
>>
File: 00026-3964655381.png (1015 KB, 832x1216)
1015 KB PNG
asuka is a what?
>>
>>109650172
https://www.youtube.com/watch?v=VuXV02VHwXY
>>
>>109659260
basedface youtubers are on anime diffusion now
damn
>>
>>109658223
Lovr
>>
>>109658023
Better cloud seevices than perchance is CivitAI
>>
>>109657059
Makes me wonder how many training hours went into Noob or Anima. I doubt it was just one blind run from 01 to 1. It's more likely they branched out, tested different possibilities, and kept evaluating the results at every finetuning step. A company putting out anime models that cover anime, /aco/, and furry content isn't going to just pompously drop one big update a year and hope for the best, that's too much risk to gamble on a single shot.
>>
>>109658223
Hagsuka
>>
>>109657697
Wife (the smol one)and daughterwife (the booba one)
>>
>>109656613
Jesus christ, she got sent flying meters out of the pool
>>
>>109658223
Nigsuka
>>
>>109658223
Asuka or just some tomboy in a wig with contacts trying to look like her?
>>
File: angery.jpg (353 KB, 1056x1536)
353 KB JPG
>>
>>109658223
Blondesuka?
>>
File: ;^(.jpg (4 KB, 124x66)
4 KB JPG
;^(
>>
>>109658223
An honorarian aryan? An honoaryan, if you will.
>>
>>109660047
Tough for you, anlet.
I've been so bored by every v5 image looking like shit that I've barely spent any this month.
>>
>>109657697
she does look cute
>>
File: Riruka-045.jpg (1.81 MB, 2496x3648)
1.81 MB JPG
>>
File: Krea2_turbo_00050_.png (2.34 MB, 1928x1088)
2.34 MB PNG
Long time no see, eh?
>>
>>109660470
>krea2 turbo
>indistinguishable from gptslop
amazing
>>
File: misc_20_61.jpg (255 KB, 832x1216)
255 KB JPG
>>109659937
For Blue Archive characters I think it's normal, if they can survive an 88mm shell to the face, they sure can send someone flying a few meters.
>>
>>109660015
Catbox?
>>
File: 1girl.jpg (370 KB, 1256x840)
370 KB JPG
https://i.4cdn.org/wsg/1787835594116293.mp4
>>
>>109660470
Fatt
>>
>>109661065
waving at viewer from front, upper body
>>
File: 1girl_2.mp4 (1015 KB, 896x576)
1015 KB
1015 KB MP4
>>109661200
highest quality, best quality, masterpiece
>>
File: image.png (69 KB, 795x254)
69 KB PNG
message by circlestone_labs, OP, posted 14 hours ago):
Even if I don't completely retrain on a new model, a proper dataset update and general upgrade to Anima would still be a huge finetune on many millions of images. This is probably at minimum $20k. A retrain on a newer, larger base model would be $100k+. I make okay money right now from the licensing thing, but not enough to immediately spend that much on a big training run when rented GPUs are more expensive than ever. Turbo and reinforcement learning R&D is basically free (I can do it on my local machine), so that is what I'm spending time on until I have the funds for a large base model pretrain.
NAI v5 recently came out, and that model was well over 1 million $ of rented compute just on the training run itself. These things are expensive if you want to make something actually good.

Uh oh! No Anima update, repeat: NO ANIMA UPDATE!!!
>>
>>109661298
Thanks for the heads up, anon, I was about to cancel my subscription that ends today because NAI5 is shit, but I guess I'll stay, local seems a dead end.
>>
>>109661298
>Turbo and reinforcement learning R&D is basically free (I can do it on my local machine), so that is what I'm spending time on until I have the funds for a large base model pretrain.
>that is what I'm spending time on until I have the funds for a large base model pretrain.
>that is what I'm spending time on
So no ControlNet, style transfer or IPA adapter...
>>
>>109661445
Making local tools was never a priority for that grifter
>>
>>109661298
If making a reasonably stable finetune costs that much, and Anima still isn't stable, what hope is there for these $50 to $100 shitmerges flooding CivitAI daily?
>>
based
>>
>>109656521
>maybe later
That would explain a LOT.

>>109656582
>It's Beksinski yes
El Basado,
>>
File: War Mace - Regen build.jpg (623 KB, 1467x2144)
623 KB JPG
>>109656651
>>
>>109661298
100k for a proper local finetune, Anima
1M for professional grade, NovelAI.
So no, your home GPU is not making out a good character or concept lora either.
Wake up local faggot.
>>
>>109661722
>>109661390
buy an ad
>>
File: file.jpg (641 KB, 768x1152)
641 KB JPG
guys my koharu is the wrong color
how fix????
>>
File: anima_is_shit.png (184 KB, 416x1371)
184 KB PNG
In case it wasn't obvious, these are the kinds of people shitting on anima. They are in here too, posting the same thing, just anonymously. That guy joined noobcord less than a month ago btw, posts almost daily about how anima is bad, broken, untrainable, deadend, dev is a grifter, etc. Like will actually just bring up the "anima is le bad" out of nowhere in an unrelated conversation.

IDK why this model in particular has attracted so many dedicated anti-fans. Even for models like Chroma, it wasn't this bad. Just be aware.
>>
>>109661730
NovelAI doesn't need an ad right now. If Anima confirms he is stopping model improvements and going the Hassaku route of turbo, aesthetic lora merges, then NovelAI becomes the ChatGPT of anime genning with literally no Claude tier competitor to challenge it. And no, PixAI isn't anywhere close to touching what NovelAI has.
>>
>>109661763
no one gave a shit about chroma because it actually garbage and no one used it.
Anima is actually a fairly good and usable model. And is absolutely "post" trainable at a purely hobbyist level(read: loras). There are valid grievances to be had but the main thing is that people are actually using it.
>>
>>109661763
He's right though. Just because it's the only anime model out there doesn't mean it deserves to be worshipped. It's been 8 months since Anima, anon, and there haven't been any significant improvements.
It's been confirmed that tdrusell won't explain how he trains his base model, because he makes money off Anima. Just be aware of that.
>>
>>109661722
Midjourney made haft a billion in 2025
NovelAI made ~5 miiilion in 2024 (lol)

The only future that awaits NSFW anime generating services is bankruptcy because jews don’t like it. Local is the only way for gooners
>>
>>109661813
Haha, enjoy those GPU rental prices. Funny how the "top name" in local anime models tdrusell himself isn't training anything of value... because he can't actually train anything of value locally. >>109661298
>>
>>109661813
2024 was the v3 era NAI 4.5 came out last year.
>>
>>109661810
>the only anime model out there
it's not tho. there was a new noob release not too long ago plus every other one down the line. krea2 is smart enough to recognize official anime styles and it's only a matter of time before we get something even better
>>
>>109661763
it's harder to train with way less margin for error, so it's much more prone to catastrophic forgetting than SDXL, which breeds resentment. Also, noobcord doesn't only hate on Anima, they shit on LAX, Chenkin, and Newbie as well. Anima just catches more heat because it sees way more active use than those models.
>>
>109661908
oh, a noobcord whiteknight iwonder who that is! what kind of loser gets kicked out of their own server for being a lolcow, and then goes to hide in noobcord?
>>
>>109661810
>It's been 8 months since Anima
It has barely been 3 months since base 1.0 was released. The model was continuously training until then.
>It's been confirmed that tdrusell won't explain how he trains his base model
WTF are you talking about? Provide a source. His training script (diffusion-pipe) is open source. He has provided an example config and a dataset for lora training best practices.
>>
>>109661953
>for lora training
>>
>>109661953
I'm talking not loras but the base model itself
>>
>>109661813
>Local is the only way for gooners
But Nvidia bought huggingface anon~
>>
>>109661960
>>109661978
fucking retards, full finetune is literally lora config, but lower the learning rate and raise the batch size a lot. it's 2 fucking knobs to adjust
>>
>>109661885
people complain that anima takes too long to gen and you think people are going to actually use krea2?
bruh
>>
>>109661757
>the nai sheen
>>
>>109660015
second, catbox please.... onegai
>>
>>109661763
Thanks for the info, but so what if Silveter came here to fud Anima? Isn't this website supposed to be about free speech? Spreading fud doesn't justify doxing someone.
>>
>>109661990
>people complain that anima takes too long to gen
where are these complaints?
>>
>>109661990
krea2 turbo is a few seconds longer than anima turbo and has full quality. big whoop
>>
>>109662107
>screenshotting posts by a pseudonymous account on a public forum
>doxxing
>>
>julien
>>
>>109660883
>>109662037
no box but here prompt:
masterpiece, best quality, (score_9, score_8:0.5), 1girl, solo, (cowboy shot), (@hata4564:0.5),
matoba risa, standing, school uniform, skirt, shirt, crossed arms, angry, nose blush, (sweat), zettai ryouiki, (thick thighs:1.4), restaurant, blurry background
>>
File: emoticon.jpg (4 KB, 300x267)
4 KB JPG
>>109661763
Alright, post some screenshots of people who finetuned Anima and are actually happy with how it turned out. :^)
>>
File: ComfyUI_00279_.png (1.76 MB, 1024x1536)
1.76 MB PNG
>>
>>109662269
lmfao
>>
>>109662269
post an example of a good, non-fucked finetune of literally any post-SDXL model (other than anima itself)
>>
>>109662135
that anyone even remotely considers using turbo like >>109662141 suggests implies that there are a lot of people who think anima takes too long to gen.
Though a large part of it is that it's shit at upscaling so you either gen lowres and only lowres or you generate at native high resolution which takes like 8x the time it takes to generate lowres on SDXL(which you can actually manually upscale without it being shit)
>>
>>109662277
catbox this please
>>
Is krea 2 able to edit images natively without nodes and loras cuckery ?
>>
>>109662448
krea2 base is for training only. I can run it but it's not as good quality
>>
>>109662472
needs an edit Lora and that only works for knowledge in the model
>>
>>109661763
>IDK why this model in particular has attracted so many dedicated anti-fans.
Anima just became frustrating for people who hoped for a model that was a complete upgrade from noob but the reality is that it is a sidegrade with pros and cons.

I know I personally had hoped for a model that had all the conceptual knowledge noob had and more, along with controlnets and tools alongside it for multi image style/vibe transfer, and multi character reference. I'm just not going to pout about not having a model like that on local yet.
>>
>>109662593
What is better in the SDXL world for ControlNet, style transfer and so on; Illustrious or NoobAI

dont troll me, i'm actually new
>>
>>109661763
I still prefer older booru models to anima because their styles are better and they do porn better. I don't need your fancy prompting, I'm good
>>
>>109662646
Oh I forgot to say for inpainting as well, which seems to be like the best way to make actual non trivial stuff
>>
>>109662646
>in the SDXL world for ControlNet, style transfer and so on; Illustrious or NoobAI
nta but I thought neither of them do controlnets and inpainting
>>
>>109662646
>>109662681
Wasn't implying that noob was better off with style transfer or controlnets so much as that those would be the bare minimum to replace some aspects like artist mixing as a method to control style. Currently Anima styles are extremely delineated. For instance, on noob I could prompt schedule and mix my way into different Yoh Yoshinari styles but attempts to do this in Anima will result in a median unless you bake a lora for each of them, not to mention if you are trying to generate a style that isn't based off of a specific artist.

That said. Noob does have controlnets for inpainting and such.
https://civitai.red/models/1376234/noobai-inpainting-controlnet
https://civitai.red/models/929685/noobai-xl-controlnet
There is even a style transfer one but I never got decent results with it
https://civitai.red/models/1233692/style-ipadapter-for-noobai-xl
>>
>>109662754
Personally I'd rather just img2img to try and prompt some changes or simply regen the same image with like 1 different tag than mess with these janky inpainting things
>>
>>109662269
Anima bross???
>>109662429
Noob,
Post Noob models: Chenkin
"Feeling lazy today" models: WAI (my beloved)
>>
>>109662885
Honestly really only used the anima one for this but the inpainting controlnetlllite for that can be used to make portrait assets with consistent character design. I don't think I have ever actually used it for true inpainting.
>>
>>109661989
Our response, anti-Anima bros?
>>
>>109661298
Knowing the dev has now switched to a "Netflix and chill" mode and only does homemade finetunes kind of demotivates.
>>
>>109662651
This Anima's for VRAMlets. If your iterations are fast, nothing's stopping you from doing txt2img, upscaling to 4k, then using context padding to nail the details. Same sharpness you'd get from a modern VAE.
One catch though: SDXL and Anima both shit the bed on backgrounds. That's where Krea 2 comes in. It runs natively at 2k, inpaints backgrounds with way more consistency, and doesn't hallucinate character details bleeding off the edge of the body like SDXL or Anima sometimes do.
>>
File: 1772039379681120.jpg (3.87 MB, 1600x2216)
3.87 MB JPG
>>109662754
Random example gens from my noob folder for what I mean. You can probably tell on some of these what artists got mixed in but overall it can create something more distinct than its parts while Anima converges. I still like Anima for a lot of reasons but autistic prompt mixing to fish out new styles is not one of them.
>>
>>109661988
Really? I mean, they're the ones gatekeeping the GPUs to begin with. Maybe they answer to the chip makers upstream, who knows, but at the end of the day they're the ones setting the value that the entire AI ecosystem revolves around.
>>
File: ComfyUI_temp_yccaq_00031_.png (2.02 MB, 1024x1536)
2.02 MB PNG
>>109662593
Anima isn't an enormous improvement over the better Illu models but overall it is absolutely a step up in terms of clarity, anatomy and composition, not to mention the upsides. Not that I have any plans of deleting my Illu models, there's still so much that hasn't been replaced yet. But to say Anima is a sidegrade is just false.
>>
>>109662885
>>109662932
What is true inpainting?
Isn't inpainting the most popular thing for complex stuff? Like with the Krita extension and stuff
>>
>>109663160
>But to say Anima is a sidegrade is just false.
Agree to disagree. Enhanced clarity without upscale and NL comprehension of some things is an enhancement but there are a lot of concepts and styles you can't milk out of Anima unless you use a lora. This makes it only an upgrade for some people and a downgrade for others. When something is six of one or half dozen of the other then I'm inclined to feel like it is a step to the side overall.
>>
File: ComfyUI_00177_.jpg (300 KB, 1056x1536)
300 KB JPG
>bocchi
>check inside
>she has a glock
>>
>>109663297
>a lot of concepts and styles you can't milk out of Anima unless you use a lora
NTA but that's true of Noob/Illu too
>>
>>109662910
>noob, chenkin, WAI
>"post-SDXL"
still waiting
>>
>>109663634
Element 7 - Void Spirit 1
>>
>>109662429
pdv7 :^)
>>
File: ComfyUI_00205_.jpg (358 KB, 1152x1536)
358 KB JPG
>>109663618
after more messing around I finally got what I intended, mostly
>>
>>109663646
>>109663710
Concession accepted.

For posterity, here are all (or most) of the failed, undertrained, schizo, or just generally borderline unusable post-SDXL finetunes:
Chroma, PonyV7, Neta, NetaLumina, Newbie, Chroma Kaleidoscope, Zeta Chroma, Nekofantasia, multiple non-public Klein 4b attempts, Sulphur, bigASP v3, Kroma (training in progress, but not looking good), Mugen (if you consider arch modifications to SDXL to be post-SDXL)

Of all these, Chroma is the closest to being decent.

Here are the successful post-SDXL finetunes:
Anima

This basic fact makes lots of people very angry. There is a consistent pattern of the community being completely unable to train anything that isn't SDXL, with tdrussel being the one exception. IMO this is the ultimate source of most of the Anima hate and FUDing. People are jealous, simple as.
>>
>>109662231
>>109662277
I want to steal your workflow, catbox it
>>
>>109663801
I will describe it for you:
- load anima base 1.0
- empty latent image 768x1024
- sample 26 steps, 4.0 CFG, euler_a, 100% denoise
- upscale latent by 1.5x, bilinear
- sample 30 steps, 55% denoise
- vae decode

that's it
>>
I don't know how to help people who ask for workflows
You know all that 90 nodes bullshit is snakeoil right?
It's just the base model + a decent style lora
You could do this on neoforge
Stop playing puzzle games in orders to generate girls standing
>>
>>109663765
But I named one. You just ignored me. What's wrong with Void Spirit?
>>
>>109664005
You have a point.
Comfy UI is not an anime model, there is no use case for use it for anime genning. The only positive thing is that it has SoTA memory managment.
Better use Swarm, Forge or Invoke
>>
>>109663622
Noob can do scat of the box
>>
>>109663160
>lolicon
>post a picture of a hag
>>
>>109661298
I don’t know why only Kobold Horde exists for LLM collabs. Why can’t there be a Diffusion Horde for fine tuning? If more than 3000 people want a new anime model, why don’t we make something like BitTorrent, but for GPUs?
>>
>>109664175
My statement remains true.
>>
>>109664181
That's Matoba Risa, she's a loli
>>
>>109664284
loli is a body type
>>
File: ComfyUI_00474_.png (1.79 MB, 1024x1536)
1.79 MB PNG
>>
>>109664382
an overdose of chocolate would be the cause of my death
>>
File: ComfyUI_00504_.png (1.74 MB, 1024x1536)
1.74 MB PNG
>>109664391
But the other 2 though
>>
>>109664315
That's Matoba Risa, she's loli-aged with a loli body-type
>>
File: le 1girl au chocolat.jpg (322 KB, 1728x1152)
322 KB JPG
>>
File: image.png (691 KB, 1111x1111)
691 KB PNG
>>109664622
>>
>>109664005
What has better inpainting tools CofmyUI or neoforge or whatever?
>>
>>109663622
Yeah. That was what I meant by this.
>When something is six of one or half dozen of the other then I'm inclined to feel like it is a step to the side overall.
But perhaps it would have been better to say they feel a bit complimentary and neither really completely replaces the other right now.
>>
File: 1780364343428256.png (1.18 MB, 1024x1024)
1.18 MB PNG
>>109661666
hnnng
saved
is it an oc?
>>
>>109664738
He's right though
>>
>>109664738
wish i had a sister who looked like that to kiss
>>
>>109653692
Is this that warhammer stuff autism?
also is that anima or noveai
she cute, wish she was my neighbor in stellaris
>>
>>109664382
>>109664498
>>109664644
>>109664738
Catbox
>>
>>109665079
ye OC, people seem to like her a lot and she's fun to gen
>genning getting better and better at NOT making extra limbs
>make a character that has extra arms and uses them
lol
>>109665247
ye 40k, it's naiv5
>>
File: button dress.png (1.16 MB, 832x1248)
1.16 MB PNG
idk how a button front dress would have a seamless lace trim on top but that's up to science to figure out
>>
>>
File: ComfyUI_00581_.png (1.76 MB, 1024x1536)
1.76 MB PNG
>>
>>109665584
Decorative buttons, there is a zipper in the back
>>
>>109666322
but it was hotter thinking about how it completely opens in front
>>
the weird thing about AI image models is that their learning curve is basically inverted compared to how humans learn to draw
>>
>>109666804
gesture anatomy perspective rendering color composition style
What I mean is that style doesn't really exist at the beginning for the human illustrator, you can't consciously decide “i have a style now” when you still don't know how to draw an arm
>>
>>109666804
>>109666822
I'd be careful saying that. We AI artists learn all those things at the same time, not in exactly reverse. Every level, like composition, anatomy, style, and lineart, is learned simultaneously.
>>
>>109666804
You could argue that is true of AI in general. The tasks AI excels at are higher order mental activities, but the kind of tasks that we have instinctive brainpower for, or are learned in early childhood (like motor skills needed for robotics) have been a lot harder to come by.
>>
>>109666804
I mean it makes sense when you remember that diffusion models "draw" in exactly the opposite way that a human does.
>>
>>109666847
But don't you think style emerges after enough individual decisions become automatic? If I generate 10,000 images over some amount of time, I will eventually stop consciously thinking about every tag and every prompt for each of them, and the way I generalize them will become recognizable, and that way of generalizing is what I call a style.
>>
>>109666893
>>109666822
Illustrators spend years acquiring deep sensorimotor skills and only later build the abstract symbolic reasoning.
>>
>>109661666
>>109656521
Simping over Warhammer 40k Total War. Once that game drops I'm gonna ditch diffusion for a good while.
>>
>>109666934
I meant more that a diffusion model, being a sophisticated guided noise removal algorithm starts with *everything*, and works down to just one thing, compared to a human who starts with nothing, and builds up.
>>
File: 00027-3598902171.png (1012 KB, 832x1216)
1012 KB PNG
you mistify me
>>
>>109661298
lmao1 million dollars for nai? someone is taking a lot of money
training novelai should not cost more than 5k lol
>>
>>109661390
AI is a MASSIVE SCAM, pic related is cheaper than NAI
>>
File: This or Anima2.jpg (264 KB, 588x872)
264 KB JPG
Anima 1 with updated dataset or a used crawler excavators??
>>
File: ComfyUI_03374_.png (1.13 MB, 1024x1024)
1.13 MB PNG
>>
>>109653692
Is this even possible to pull off locally? Like in Anima?
>>
>>109667270
I've been using Nai since early v3 days, but I'd imagine you'd need a multiarm lora to better control them, whichever style you're going for (this is just momisan), your standard quality tags and
>action, battle, holding staff, looking away
Someone with local knowledge can absolutely steer you better than me lol
>>109666955
Enjoy when it hits!
>>
Getting a little sick of Gradio/Forge UI being a slow piece of shit, and I'm a hyper autistic motherfucker when it comes to efficiency/personal quality of life. I started writing a Lisp program to output prompts and stuff and slightly modified the "prompts_from_file" script to just load the file from a path each generation. Evidently, I don't really want to spend a lot of time working on this program (even if it would, in theory, give me the exact template functionality I've wanted to be able to hotswap outfits on the fly without bloating the prompt with stuff that affects composition like footwear) so I guess what I'm asking is if the WebUI for SD is a slow piece of shit for everyone or if I've managed to fuck up my shit with a dumbass amount of extensions.
>>
>>109667640
Hekin’ cute
>>
>>109667640
>Extra Armor
That rule can now be on non-vehicle units? What does it do?
>>
>>109667640
Your gens are very cool anonnie, I enjoy them.
>>109668059
I don't know what you're calling slow, but my Forge Neo works fine. Also curious about the your clothes swap on the fly mechanism.
>>109666965
Why Nazi and why brown?
>>
>>109667147
Why?
>>
We need an /adt/ cord ASAP
>>
File: cord.jpg (233 KB, 832x1024)
233 KB JPG
>>
>>109669342
Exelent now add an adt to that cord
>>
>>109669451
data Nat  = Zero | Succ Nat
>>
>>109669451
Anima can't do it...
>>
>>109669533
Oh no.... a-nd K-krea..?
>>
>>109669620
Neither, ControlNet or nothing
>>
Anime edit model. I'm not picky :(
>>
>>109669080
?
>>
File: qtpie.jpg (444 KB, 1344x1536)
444 KB JPG
>>
>>
>>109669087
ani made a discord for us all to keep the schizos out
>>
>>109668690
Extra Infantry Armor and Extra Vehicle Armor are separate values, but I wonder if she grants both when holding on to his back lol
>>
Danbooru 1024 40K LoRA for z-image (base or turbo, training on base) at step 1000. LR is 0.00001, keeping it low because there are many concepts it has to learn from the dataset., so I do expect it to keep converging until 12000 or so.
Even at step 1000 though, the prompted for "micro-bikini" is beginning to appear, so it's a good sign,
>>
>>109670741
very interesting
cautiously optimistic
>>
>>109670777
Yeah I'm hoping it doesn't collapse or forget, it would be nice to have a fast little thing like z-image turbo for NSFW anime. It probably WILL bias things towards big boobs and excess bodily fluids, though, since Danbooru 1024 is full of that.
>>
>>109669967
Oh mahiro hi res fix! :3 hello mahiro~
>>109670626
Where tho?
>>109670741
Neat, did you test Illustrious Z Turbo Anime? It's closed source, apparently they trained it with a huge dataset of anime, but even so there are problems with character learning and the anatomy is all broken.
>>
>>109670626
>>109670626
>>109671139
kys and maybe stay in discord with the rest of you trannies, 4chan is for cisgenders
>>
>>109671139
>It's closed source,
Yplou can try three (3) gens for free with a burner in illustrious webpage.
Spoiler: is shit.
>>
>>109671169
>projecting your discord faggotry onto others
The only /g/eneral that has a discord is /ldg/. No anime general has a discord.
>>
Holy shit, NAI5 feels like doing acupuncture, no seed instability at all.
>>
>>109667640
Cool, this is txt2img without inpainting?
>>
>>109671435
example?
>>
>>109671488
You have to believe me anon!
Imagine doing batches of seeds(with autoclick), and the style changes drastically. For example, with anime coloring, official art and high complexity tag as the only stylistic tags, I have a chance of getting a painterly gen, in 4 gens, and the face of the character shifting of style various times!
>>
I like the idea of Ani fudding NovelAI
>>
File: Zero hesitation.jpg (816 KB, 1024x1400)
816 KB JPG
>>109671444
I skipped the RNG and just photoshopped the text lol
>>
>>109667147
cute
>>
>>109665584
nice
>>
No....engagin....:(
>>
>>109671989
adorbs but unsettling!
>>
>>109669342
kawaii phone holding rin
>>
>>
>>109656662
formidably sexy
>>
>>109655481
kawaii
>>
>>109655314
Kushida is so sexy!
>>
>>109672486
Strong girl. I need a strong anime girlfriend to protect me.
>>
Yes engaging...:)
>>
>>109672486
That's a mahou shoujo, no regular shoujo is that strong.
>>
File: 1776794053751112.jpg (631 KB, 1024x1024)
631 KB JPG
>>
>>109654443
nice girl
>>
>>
>>
>>
File: file.png (2.43 MB, 1216x832)
2.43 MB PNG
>>
>>109672769
Very nice.
>>
Migrate when ready.
>>109672835
>>109672835
>>109672835
>>
>>109672836
thanks for bakering
>>
>>109671176
>Spoiler: is shit.
at least one is decent, indeed better than the other two (allow LORAs character and styles), shame nvidia will nuke it.
>>
>>109669074
There's an active latency to type into the prompt field. It's almost definitely related to autocomplete, to be honest, but if I'm putting the effort in, I might as well also implement the autocomplete in my editor of choice eventually.

My clothes mechanism is pretty simple; You can pretty easily separate most 'outfits' into different parts of the body. Headwear, Torso, Legwear, Footwear, sometimes extras or accessories. I think it would be cool to do some slightly more advanced logic based on the tags apparent in the prompt, but the basic idea isn't far off from just variables.

You add a "tag" that sets the outfit, and you include corresponding "tags" for whether the image should include the headwear/torso/legs/feet, and the outputted prompt will automatically exclude parts of the outfit based on whether the tag for the part of the body is present.

It's basically just a more advanced wildcard mechanism. I really tried to finagle with the existing wildcard/dynamic prompt extension, but the best I was able to achieve was including the outfit parts individually with the full namespace.

Basically, currently I have to do "__arcueid/remake/base__, __arcueid/remake/torso__, __arcueid/remake/legs__, __arcueid/remake/footwear__". In my more idealized system (which is now written in Lisp, but it basically just means I don't have to repeat 'arcueid/remake' a ton), it'd be written
(prompt
...some other tags...
(base "arcueid/remake") :torso :legs :footwear
...the rest of the tags...)

I'll probably include a mechanism for setting the outfit to someone else's outfit while more fundamental details like the face are set to the appropriate character for stuff like cosplay. Theoretically, body types and proportions could probably be set (and overwritten) as well. I'm not sure I'll ever really publish it due to the niche nature of it being written in Common Lisp, but maybe it sounds like a cool idea to someone.



[Advertise on 4chan]

Delete Post: [File Only] Style:
[Disable Mobile View / Use Desktop Site]

[Enable Mobile View / Use Mobile Site]

All trademarks and copyrights on this page are owned by their respective parties. Images uploaded are the responsibility of the Poster. Comments are owned by the Poster.