[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_148.webm (2.31 MB, 1286x1112)
2.31 MB
2.31 MB WEBM
Previous: >>109879682

>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: included above

>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.jpgcollage >>106520607

>/adt/ logo:
https://files.catbox.moe/pakfgt.pngcollage
>>
File: Krea2_turbo_02571_.jpg (1.63 MB, 1776x2368)
1.63 MB JPG
>>
>>109932045
Whoa buster, forgot your

>Beware inorganic activity
https://rentry.org/animanon
>>
>>109932055
>>109932057
Off topic
>>
>>109932045
Cool collage
>>
>>109932271
You won't and can't do shit pussy
>>
>>109932045
Animated collage stronkest
>>
>>
>>109932571
Devious pits
>>
File: ComfyUI_temp_rhhav_00005_.png (3.05 MB, 1536x1024)
3.05 MB PNG
>>109932571
Nice Rall
>>
File: ComfyUI_temp_jaijz_00022_.png (2.23 MB, 1024x1536)
2.23 MB PNG
Imma just Nanapost idgaf
>>
>>
File: 356272727.jpg (1.1 MB, 1472x2176)
1.1 MB JPG
>>
File: 1787465351792.jpg (422 KB, 1080x1349)
422 KB JPG
Do you all have an OC that you test new models with? I find using characters from IPs forces certain artstyles a lot
>>
File: Krea2_turbo_01640_.png (1.05 MB, 1184x888)
1.05 MB PNG
>>
File: edi.png (2.29 MB, 1184x1472)
2.29 MB PNG
>>109934745
I'm hungry
>>
>>109934868
Pretty hot ngl, I didn't think to try darker skintones with her. Might want to make the hair blonde or white tho
>>
File: 1787729801944796.png (2.91 MB, 1344x1920)
2.91 MB PNG
>look up what UIs to use for video
>comfyUI pops up first
what the fuck anon actually made it big, its a full fledged product now
>>
File: 62617161717.jpg (1.15 MB, 2304x1408)
1.15 MB JPG
>>
File: 9586784583.png (3.12 MB, 1150x1700)
3.12 MB PNG
>>
>>109935137
The ui went to shit because of corpo interests. Welcome to the anistudio waiting room
>>
File: 7878.png (1.95 MB, 990x1467)
1.95 MB PNG
>>109935262
>>
File: Buggos.jpg (1.72 MB, 2048x1536)
1.72 MB JPG
>>
File: ComfyUI_07550_.png (1.68 MB, 1920x1080)
1.68 MB PNG
>>
I've been out of the loop for a while (couple of years), is it still better to run the base model and then add one or two Loras at most or have we gotten to the point where you can stack Loras on top of some random checkpoint like a maniac without images melting? Mostly interested in Anima, and there's like a few thousand distills available

>>109935262
Very neat style, looks like an official gacha splash image
>>
File: axe.png (1.45 MB, 1920x1080)
1.45 MB PNG
>>
File: witches (1).jpg (876 KB, 2483x1241)
876 KB JPG
I wonder if Blair witch has a big hat too.
>>
>>109932045
>ranfaggot lost
Based
>>
https://huggingface.co/Noctaluna/Noct-Q-Anime-Uncensored-Qwen-Image-2.1?not-for-all-audiences=true
I don't have a gpu but I came across this.
>>
>>109935262
How did you prompt this?
>>
File: ump45.png (2.41 MB, 1024x1536)
2.41 MB PNG
>>109937655
Not OC but you can do it easily using this lora
https://civitai.red/models/2900619/gacha-splash-art?modelVersionId=3279781
>>
File: Krea2_turbo_00470_.png (1.58 MB, 1448x1448)
1.58 MB PNG
See that "frozen windows screen" background? Is that some sort of halfway refusal fuckup from krea2? I see that in about 1/8 gens. I think also it might be a LoRA thing. Ideas?
>>
>>109937810
do you really need a lora for it? i always just used fading white borders in my prompt
>>
File: AniStudio-01026.png (1.08 MB, 896x1408)
1.08 MB PNG
>>109937842
it normally happens on white backgrounds for me but I still get it on sfw prompts so must be something about how white backgrounds/alpha were trained rather than refusals. might have been confused when converting rgba->rgb and nobody caught it in time. never happens when I include a background
>>
>>109938929
Pedo
>>
>>109939103
>sees drawing of child
>immediately thinks of sex
>Feels shame and projects his feelings onto others
>>
>>109939274
But
>>109932057
Shows his activities on other boards though
He can't get a job for good reason
>>
>>109939395
Go away raped schizo
>>
>>109939436
How is the truth schizophrenic?
Can you disprove these logs?
>>
>>109939577
No we purposely waited before showing those post for good reason he has a 4+ year history and has been caught posting and defending that material in the past.
We don't false flag and let the lolcow dig his own grave. He's well past the point of return anyways.
>>
File: Bnnuyangles.jpg (326 KB, 1536x1024)
326 KB JPG
>>
File: 1784020771173236.gif (600 KB, 350x200)
600 KB GIF
>>109932045
I love Illustrious because the result are varied everytime i hit "Generate"

Anima cant do this shit. Thats why its frustating to use it
>>
>>109939728
>tfw still use illustrious for Precise edit
>>
>>109939601
>We don't false flag
Yes you do
>>
>>109939601
Nothing says "We are the mentally stable ones" like stalking some literal who on an anime website over multiple years and just casually admitting it, top kek

>he posted some legal images I don't like, I am fully justified in wasting every waking second of my life on this person!
You're even more lulzcow than trani is. For any normal person that would take some effort, but it's natural to you, isn't it.
>>
>>109940097
>>109939983
You don't have anything to offer other than arguing, we separated ourselves from him multiple times and he inserts himself within these spaces. He abused this general to shill his software only to not deliver and anons turned on him.
If you can't understand this I don't know what else to say other than leave me alone.
>>
is julien here? anon said i should post a github ticket with the question but i would like to ask here
>>
>>109940154
Who is your male elephant? Human name?
>>
>>109940132
Has anyone ever thanked you for your great dedication to keeping this board clean from some dumb shitposter? No. Never. Everyone is, at the very least, annoyed by you, you absolute failure.
I don't like Ani, I don't support Ani, but you being unable to let this go, unable to contain your spaghetti when you see him post one of his 10 billion identical mascot catgirls is unsightly.

Just leave. Try to do something with your life
>>
>>109940154
Post a GitHub ticket instead :)
>>
>>109940187
>I don't like Ani, I don't support Ani
Weird. There isn't any reason to dislike him but you do have a point about the schizo being annoying
>>
>>109940187
Why do you think it's only one person?
Are you unable to accept Ani's biggest opposition is himself and how he interacts with online spaces?
This isn't the only online space he's been kicked out of either, it's multiple including other diffusion generals.
>>
>>109940234
I've never had anything to do with Ani. Some of the proof they gathered may even be correct for all I know. I don't hate Ani, I don't like Ani, I just don't really give a shit.
Whatever Ani did is just a billion times less annoying to me than this one constantly malding about him, whether he's here or not, whether he did anything or not.
>>
>>109940321
So? Get over it. You're the bigger lulzcow that you keep malding about it.
>>
>>109940259
>Why do you think it's only one person?
I've explained this before. This thread doesn't get much traffic, only a handful of IPs. The likelihood of there being multiple people as mentally ill as you within such a small cohort just isn't likely.
>>
This is brittbong tier of damage control.
>>
>>109940259
It's your obsession. You keep posting about him. Let it go, the problem goes away. None of this affects you in the real world, right? Who cares about Anistudio? Who cares how many stars it has?
Just let it go.
>>
>>109940461
Based schizo anon
>>
>>109940461
Absolute fucking lulzcow.
At the very least, remember these words the next time someone calls you out, before you spout your self-righteous "I am only doing this thread a service!" shit.
>>
>>109939726
Krea could never
>>
>>109940530
it is krea
>>
im not gay therefore i do not support avatarfags simpleas
>>
>>109940534
>lying
Krea does not know my wife to put her in that outfit.
>>
>>109940579
loras do newfag
>>
>>109940587
Lying gets you nowhere man. It's okay, $25 gens are acceptable here
>>
>>109940461
Alright, man, but please, have some joy in your life, OK?
>>
why is julien so bad at false flagging?
>>
>>109940507
>>109940461
>lose argument
>samefag off the cuff schizo post
This doesn't work
>>109940654
He's beyond help and can't reflect on his actions which is why comfy abandoned him.
>>
has anyone trained style loras for anima on civitai? did it turn out well?
I don't wanna do it local
>>
>>109940745
>anima
>>
File: ComfyUI_07681_.png (1.06 MB, 1024x1024)
1.06 MB PNG
>>
File: ComfyUI_temp_qhzyb_00076_.png (2.96 MB, 1024x1536)
2.96 MB PNG
>>109940745
As I've said in previous posts it's more stubborn to train for than Illu but yes you can train loras for it just fine on civitai
>>
Holy fucking shit. This Debotard literally on every single /g/ thread.
>>
I think bullying ani is just /g/ culture now and it spreads to other boards
>>
File: Animal Fight.jpg (391 KB, 1792x1792)
391 KB JPG
>>
>>
>>
hi
>>
File: 234235253.png (1.37 MB, 1920x1080)
1.37 MB PNG
>>
File: 8243092.png (1.68 MB, 1920x1080)
1.68 MB PNG
>>
Which version of Anima do you people use? I am generally getting better results using Turbo, at 1 or 1.2-1.5 CFG than on Base, but I've yet to try WAI-Anima or any of the other finetunes.
>>
>>109943649
I don't use anima. I use I'll for nsfw and krea for sfw
>>
File: tora.png (1.46 MB, 1920x1088)
1.46 MB PNG
>>
>>109943649
Turbo at 1 CFG, Euler.
>>
>>109942414
she scares me
>>109943649
base
>>
File: horny miku.jpg (371 KB, 768x1152)
371 KB JPG
>>
File: 2.jpg (400 KB, 768x1152)
400 KB JPG
>>109944917
>>
File: Tard wrangling v5.jpg (196 KB, 832x1216)
196 KB JPG
>>
>>109945357
are you saying she's stupid
>>
>>109942779
nice
>>
Is krea 2 base worth using over the turbo?
>>
>>109946429
no. it's for training only
>>
>>109946547
Doesn't the base have more seed variance and better prompt adherence? that's what i remember from flux 2 klein base vs turbo.
>>
>>109946599
more variation yeah but the quality is worse and it's the same llm so there isn't much difference for adherence
>>
>>109946607
it should tho, because higher cfg?
>>
>>109946616
I guess because of negs as well but NAG exists. it also takes much longer to inference. you will see what I mean about the quality if you are bothering at all with trying it out.
>>
>>109946629
I did try, the turbo takes like 30s, while the raw like 300s.
>>
File: 1790780485209164.png (1.96 MB, 1216x832)
1.96 MB PNG
>>
File: Dat HL3 tho.jpg (316 KB, 832x1216)
316 KB JPG
>>109946209
kek
I MEAN...
>>
>>109947400
Amazing NAI gen SAAR!
>>
File: AniStudio-00268.webm (2.72 MB, 960x640)
2.72 MB
2.72 MB WEBM
>>
>>109947962
cute, requesting sanae version
>>
>>109947962
Did you really post death threats online? Wtf julien
>>
>>109948094
Fraid xe did
Towards a coworker, even
No wonder xe's unemployable and living off daddy's money like a leech
>>
>>109947995
>>109948094
>>109948184
>immediate
You really are just here all day aren't you
>>
Posting a few before I pass out
https://files.catbox.moe/3escao.jpg
>>
File: AniStudio-05008.png (1.07 MB, 1088x832)
1.07 MB PNG
>>109948087
I will try after the video batch is done. I am hoping krea knows her well enough.

>>109948211
qt 3.14
>>
>>109948262
>not knowing 2hus
i'd be shocked but it does happen
>>
>>109948262
Can you answer the question? I'm also curious because if true that's really not okay >>109948094
>>
File: AniStudio-05246.png (905 KB, 1280x896)
905 KB PNG
>>109948087
I suppose this is close enough. trying to find a seed where she is immobilised and then I can animate
>>
>>109948643
oh shit thanks lol
>>
>>109948643
turbovirgin kek
>>
>>109948610
He's ignoring you, you cuck. Enough bleating.
>>
>>109948676
He can't score pussy or bussy.
You have to feel bad for a middle aged man that ERPs with other middle aged men
>>
>>109948665
https://files.catbox.moe/7a48yb.mp4
there was a nip slip so here it is, simlish va and all. I am waiting for one I can actually post for the collage maybe
>>
>>109948936
chef's kiss, thank you
>>
>>109948976
>>109948987
Off topic, you want >>>/g/ldg
>>
>>109948991
sowy :3
>>
File: AniStudio-00275.mp4 (2.81 MB, 960x640)
2.81 MB
2.81 MB MP4
>>109948936
>>109948961
bonus and won't get the collage b& ;)
>>
>>109949175
hnnnnng
great job, thank you!
>>
File: Krea2_turbo_01727_.png (828 KB, 1024x1024)
828 KB PNG
>>
File: AniStudio-05268.png (2.22 MB, 1024x1536)
2.22 MB PNG
>>
is there like a VLM where i give it an image of a character and it tells me which danbooru artists i can use to mimic the style? maybe the WD tagger could work but surely theres something newer by now
>>
>>109951709
for tags only there is joycaption. no idea if anything after that was ever improved. dunno what people use for nlp but it's probably a local model with a decent system prompt
>>
Does anyone have a better way of creating PC-98 pixel art? My current workflow is to take a regular illustration and doing post processing to reduce its resolution, limiting its colors, and transforming it to pixel art. This has given me better results than just genning pixel art and fixing it to become actual pixel art. Pic related is an example that I was able to make.
>>
>>109951804
I've tried these 3 and had the best results with pixydust.
https://github.com/jenissimo/unfake.js/
https://github.com/sousakujikken/ComfyUI-PixydustQuantizer
https://github.com/Hugo-Dz/spritefusion-pixel-snapper
>>
>>109951804
nice. don't have much experience with it myself
>>
>>
>>109953711
Who's on left?
>>
>>109953711
nice
>>
>>109944917
cute miku
>>
>>109953711
make more of this
>>
Local bros, how are we doing?
>>
>>109953711
Is that krea or NAI V5?
>>
File: Le_Malin.jpg (1.12 MB, 3072x3072)
1.12 MB JPG
>>
File: ComfyUI_temp_nufgn_00094_.png (2.78 MB, 1024x1536)
2.78 MB PNG
>>109955813
Great. Krea is a lot of fun
>>
File: MegucaFistWizard.jpg (1.9 MB, 1024x3072)
1.9 MB JPG
>>109953763
OC
>>109956303
V5
>>109955500
Will do, lol, although this is a v4.5, I'll see if I can remake it later
>>
This thread is fun and cute and sexy!
>>
>>109956929
Local lost so bad, the best thing we have is anima which is too small.
>>
>>109957046
Krea2 has been crushing it recently so idk what you are on about
>>
Is julien here? I have an important question
>>
>>109957052
Doesn't count, since there is no proper anime tune, so you need a lot of loras. That's why anima is the best we currently have.
>>
>>109957052
If you enjoy genning Miku and nameless black haired girls, and enjoy training a lora for every single style - maybe.
>>
>>109957052
Krea2 does not know my favorite artists, or my favorite tags
A lora is not going to fix that
>>
>>109957068
But anima also requires loras so why use the dumber model?
>>
>>109957081
It needs way less loras tho since it know a lot of stuff by default.
>>
File: ComfyUI_temp_nufgn_00089_.png (2.95 MB, 1024x1536)
2.95 MB PNG
>>109957068
I need fewer loras for Krea than Illu/Anima because its prompt adherence is so good you can just build characters with prompts. It doesn't really know Terra, for example. Not that there's anything wrong with loras in the first place.

>>109957078
>A lora is not going to fix that
What do you mean? A lora is exactly the way you fix that. Are you a bit stupid?
>>
sigh
>>
>>109957105
It also cant do nsfw like anima. Idk if NAI v5 can, the only nai model i used was the leaked one.
>>
File: Paca.jpg (552 KB, 1792x1792)
552 KB JPG
>>
>>109957121
It easily does NFSW, that claim has been BTFO in this thread multiple times
>>
>>109957150
Doesn't it need the filter bypass node or lora tho?
>>
>>109957105
Are we regressing back to stable diffusion days? A trained model that already knows the concepts outperforms a model coupled with a stack of loras.
>>
>>109957169
Exactly that!
>>
>>109957163
A 26 MB file or an 'unlocked' checkpoint which is pretty much all of them

>>109957169
Exactly why Krea is better, it knows way more stuff innately. As I said I have fewer loras for it than for other models
>>
>>109957169
This shit starts every time a promising local model without dan finetune appears. Guys are posting nameless 1girl standing and screech "LORA LORA LORA" at anyone who says that danbooru finetune is very desirable.
Every. Fucking. Time.
>>
>>109957190
wasn't there a delta thing for comfy that fixed the safety stuff without a lora tho?
>>
>>109957222
Possibly. Comfy's got all sorts of gadgets I still haven't tried yet
>>
>update cumfart ui
>speed halved and 2 of my workflows don't work anymore
i hate the fennec faggot so much bros
>>
>>109957204
It's why NAI is still king.
Krea IS good though, but it's too much work, not a lot of support compared to anima, and not a lot of help for those interested.
>>
>>109957249
All the tunes and the majority of loras for krea 2 are photoreal slop for whatever reason.
>>
>>109957316
Where is your loras?
>>
>>109957316
There are anime tunes for it now.
>>
>>109957316
Imagine looking at 3dpd.
>>
>>109956929
Why did she cast flammable fart instead of fist?
>>
>>109957341
I wouldn't call training on 200k images a proper tune tho. And are you talking about the wulver one?
>>
>>109957403
That would be you moving the goalposts, I'm just saying not all tunes are realismslop. Besides, the all-purpose ones work perfectly fine for anime.
>>
>>109957418
Like what? the loadestones one or?
>>
>>109957466
You seem to keep naming some. I like DaSiWa.
>>
>>109957617
There are 4 versions of it...
>>
>280+ posts
>Anima still can't gen Cabba from DBS
what the fuck
>>
>>109957647
The Cute Disaster one is the only one I've tried
>>
>>109957617
>>109958384
got a link? first time im hearing of this one. ive been trying the kroma 0.1 lora but its pretty meh
>>
>Kroma
>several months training on four RTX 6000 Pros on a subset of the Chroma dataset
>barely capable of the most basic NSFW, anatomy is a mess, massive degradation and forgetting of Krea's base knowledge
>Wulver
>larger anime/furry dataset than even Kroma
>$20k of cloud compute spent so far (literally more than anima preview1)
>it's learned basically fucking nothing
the absolute state
>>
>>109959009
Wulver has similar size dataset to anima, apparently it's 1.5M images.
>>
File: Roko1.jpg (1.63 MB, 1328x1856)
1.63 MB JPG
>>
Tested wulver, and holy shit that model is ass. It somehow has worse anime knowledge than base krea 2.
>>
so I went and trained the artist lora I wanted on civitai.
90 images, I let the auto tagger do the job, deleted some tags that I were wrong, then it trained 10 epochs.
the result is ... generally correct but kinda weird.
epoch 7 is closest to the actual style. after that maybe 6 or 8. epoch 10 doesn't resemble the input style at all. is this normal?
next is - and this seems to be specific to epoch 7 - it has incorrectly learned to make the background of gens very dark. now there are quite many pics with dark background in the dataset so I'm thinking either:
* the background wasn't tagged correctly so the model learnt it as part of the "style" instead of ignoring it
* there were just too many so there is bias (is it better to have less but unique pics in the dataset?)
the worst part is that since I last checked someone else ALREADY MADE a lora of this exact artist and theirs is just so much better. still an interesting experiment I guess.
>>
>>109959699
anima btw
>>
File: Roko2.jpg (1.73 MB, 1328x1856)
1.73 MB JPG
>>
Can i request some tamamo gens?
>>
File: ComfyUI_temp_yunjh_00107_.png (1.61 MB, 1368x1024)
1.61 MB PNG
like this?
>>
>>109957373
>POMF
>Uh oh, stinky!
>>
File: ComfyUI_temp_yunjh_00114_.png (1.79 MB, 1024x1432)
1.79 MB PNG
or like this?
>>
>>109960464
No, i like my mature tamamo thank you. First one better.
>>
File: ComfyUI_temp_yunjh_00124_.png (3.36 MB, 2048x1536)
3.36 MB PNG
>>109960482
>>
File: ComfyUI_temp_yunjh_00128_.png (2.17 MB, 1512x1024)
2.17 MB PNG
>>
File: tamago.jpg (194 KB, 832x1216)
194 KB JPG
>>109960331
>>>/e/3141440
>>
>>109960544
whoa..
>>
>>109960512
Ty annon, you made my day better!
>>
File: ComfyUI_temp_yunjh_00137_.png (2.54 MB, 1368x1024)
2.54 MB PNG
>>109960558
I like her too but nice
>>
>>109960569
What model is that, i love the textures and lighting.
>>
File: ComfyUI_temp_yunjh_00139_.png (2.06 MB, 1024x1320)
2.06 MB PNG
>>
>>109960581
You must be on mobile since you can't see filenames, it's says Comfy
>>
File: tamamo.jpg (490 KB, 2025x1733)
490 KB JPG
>>109960581
catbox is being fucky but it's
>>
>>109950987
moar like this
>>
File: ComfyUI_temp_yunjh_00148_.png (2.77 MB, 1776x1328)
2.77 MB PNG
>>
>>109960617
No way that's sdxl... i m surprised that such an old model can look good.
>>
>>109960603
You can use api models in comfy tho, that's why i was asking.
>>
>>109960657
it was with the/a niji lora as well
I find it quite nice
>>
File: ComfyUI_temp_yunjh_00155_.png (1.98 MB, 1024x1320)
1.98 MB PNG
>>
File: ComfyUI_temp_yunjh_00163_.png (582 KB, 1368x1024)
582 KB PNG
>>
File: ComfyUI_temp_dporq_00052_.png (3.94 MB, 1843x1382)
3.94 MB PNG
>>
>>109960617
>catbox is being fucky
Every time I go to upload cunny for bros
Every
Damn
Time
>>
>>109960717
>Tamago9 anomalies in the background rocking out
>>
>>
>>
>>
>>
>>109957316
even realistic loras seemed to translate the general concept into anime pretty well the few times iv tried it.
>>
>>
File: Muscle memory.jpg (204 KB, 832x1216)
204 KB JPG
>>
>>109960544
she is made of plastic
>>
>>109962614
>photo (medium), figure, photorealistic
I MEAN...
>>
>>109960544
You should make more
>>
>>109962614
In anime, they are made of cardboard, can't imagine that being any better!
>>
So excited that her game is finally coming out in January. It's been such a long wait...
>>
>>109962898
Plz do tell me more?
>>
If you have many prompts you want to try but can't keep track, this app generates all the possible combinations for you and a checklist.

Oh catbox doesn't allow html files, nvm
>>
https://21stcenturycoomer.github.io/Prompt-combination/
>>
File: umumikon.png (3.16 MB, 1824x1248)
3.16 MB PNG
>>109963613
Fate/Extra Record is releasing in Jan 2026, after years of development hell. It's a remake of the game that introduced Nero and Tamamo. Except with non-terrible gameplay this time.
>>
File: ComfyUI_00025.png (2.66 MB, 2048x2048)
2.66 MB PNG
Ming-image-0.1 isn't bad for generic anime.
>>
File: ComfyUI_00037.jpg (1.18 MB, 2048x2048)
1.18 MB JPG
Here I prompted ming to create a 3x3 grid of different emotions. For a small model, at the moment, it takes up a LOT of memory, I see it using 40 GB due to the default workflow using qwen 3.8 27b. That needs to be optimized.
>>
>>109964492
yeah thats way too beefy for my laptop, doubt ill even be able to run ideogram 4.5 when its out
>>
File: 4.png (261 KB, 512x512)
261 KB PNG
mfw 8 GB VRAM
>>
>>109964584
It's only a 6B model, and it probably does not need qwen 3.8 27b to enhance the prompt. They'll optimize it soon I think.
>>
>>109964705
The text encoder is also quite large.
>>
File: ComfyUI_00041.jpg (880 KB, 2728x1536)
880 KB JPG
>>109964787
Yeah I think it's a tradeoff for it being able to do stuff like magazine layouts. It does, however, gen 2048x2048 as its preferred resolution. No more upscaling needed for huge gens.
>>
File: ComfyUI_00044.jpg (821 KB, 2048x2048)
821 KB JPG
last ming-image-0.1 for now, I think it gives the general idea of what it can do
>>
>>109964492
Did you describe each emotion separately?
>>
>>109965049
It's pretty neat, but really not as good at text, as I would've hoped for.
>>
>>109965133
A set of nine images depicting a chibi bratty 12-year-old anime girl. She has long, shiny, saturated sky-blue hair with tapered wispy flyaway strands — anime hair vents — that break the outline of her face. Her hair catches the light with bright highlights. Her eyes are a vivid blue with long lashes; her eyelids droop down to touch the tops of her irises and lift to touch the bottoms, giving her a half-lidded, sly look. Her eyes have visible eyelashes at the top corners. 

She wears a dark navy blue beret which has a white circular buttom with a colorful Google "G" emblem on the right side, tilted on her head and a crisp white long-sleeved blouse with a sailor collar and a visible button placket, a dark navy blue bow at the chest, and thin suspenders running over the shoulders attached to her skirt with small skinny diamond-shaped brass buttons on either side. Her skirt is a pleated navy blue with a white stripe near the hem. She wears white thigh-high socks and black mary jane shoes.

Each girl is depicted as a nendroid figurine in side view with stand and holder, and shows one of the following emotions: shy, curious, happy, pouting, flustered, angry, surprised, sad, sleepy


Just a list basically
>>
>>109964201
cute aigold
>>
>>
>>
File: ComfyUI_temp_afqjl_00047_.png (1.45 MB, 1024x1536)
1.45 MB PNG
>>109959699
Your experience isn't unusual. It isn't always the last epochs that look best. And yes you must tag things to "separate them out" so it's possible that wasn't done sufficiently though that only works to a certain extent if a characteristic is just so dominant.

Try throwing the other better lora into this and see if you can see the training data
https://xypher7.github.io/lora-metadata-viewer/

In general I would say you want more steps at a slower learning rate for Anima. And more epochs to choose from because the results tend to be quite erratic.
>>
File: left other, right mine.png (99 KB, 1851x416)
99 KB PNG
>>109966416
thanks
the lora viewer is interesting, but I have no idea what this means for me.
I did not change the civitai default training settings btw.
the tag frequency is visible and I guess the other lora had better/more comprehensive tagging.
>>
File: ComfyUI_temp_jazpo_00005_.png (2.95 MB, 1024x1536)
2.95 MB PNG
>>109966921
Tagging wouldn't affect it that much. The main things that jump out at me are
1. As I mentioned before they use a slower learning rate (0.00004 vs you used 0.0001)
2. CAME optimizer vs. Adam (but civ doesn't have CAME so I guess just ignore that, Adam is fine)
3. Cosine w/ restarts scheduler vs constant (I prefer the former, though constant is fine)
4. You only used 1500 steps against 90 images, which means each image doesn't actually get "seen" that many times. I often use double that number of steps with Anima (don't compare the step counts, they mean different things with different trainers)
5. 90 images suggests to me you maybe were a bit unfocused in what images you chose, better to have 45 coherent well-chosen images than 90 okay ones
>>
>>
File: 1790972996822.png (2.78 MB, 1472x2112)
2.78 MB PNG
Mirko-san
>>
File: AniStudio-01443.png (2.46 MB, 1024x1536)
2.46 MB PNG
>>109960631
>>
File: old one.jpg (351 KB, 1152x1536)
351 KB JPG
>>
>>
File: ComfyUI_temp_yzkht_00022_.png (3.32 MB, 1536x1024)
3.32 MB PNG
>>
File: I still have 2 HP.jpg (958 KB, 1467x2144)
958 KB JPG
https://n.uguu.se/oJWRqnBi.jpg
>>
For anima does gening at 1536x1536 and then upscalling to 2048x2048 work?
>>
>>109969483
no it doesn't do that high of resolution well
>>
how do you prompt anima for soles in thick fabric socks/stockings/thighhighs? too much barefoot and transparent legwear everywhere, not enough people of true culture
>>
>>109969497
animayume 1.5 handles 1536x1536 just fine, even native 2048x2048 works fine.
>>
Post rare Depsis.
>>
>>
>>109969464
File went 404. Pls reupload.
>>
>>109971138
https://h.uguu.se/ksKqJcXV.jpg
Catbox not working for me is crampin my style!
>>
>>
>>109971210
Thanks for the reup.
>>
>>109955500
>>
File: animaTurboINT8.jpg (834 KB, 1024x1024)
834 KB JPG
>>
>>109971400
>anima turbo int8
whats that
>>
>>109971584
literally anima-turbo but int8 quant
https://civitai.red/models/2760544/anima-turbo-int8
Anima model, turbo variant (the one that generates at 12 steps CFG 1.0 instead of 30, and you can't use negative prompt because of cfg 1.0) and int8 that turns it into almost half size smaller just ~2GB model (+text encoder and vae)
slightly worse quality than normal anima-turbo but great for quickly trying stuff out on my 4GB vram gpu, probably useless if you have actually good modern gpu
>>
>>109970580
That’s a cool one
>>
>>
File: cute girl.jpg (368 KB, 896x1216)
368 KB JPG
>>109971676
I tested anima-base as INT8 (there's one on huggingface) and it turns out it's just slower than the normal one on my hardware
>>
>>109972429
Catbox please
>>
>>
>>
>>109972958
pretty nice, what model did you use
>>
>>109972830
Nova Anime XL, no workflow cuz I'm a cloudshitter, I didn't save the prompt but it was something close to
luse maonang, (shiosato jun:0.3), mylene rafa holfort, 1girl, solo, office lady, black hairband, bow hairband, black ribbon, silver hair, (blue eyes:0.5), white shirt, partially unbuttoned, large breasts, (huge breasts:0.2), black skirt, waving, standing
Negative: id card, lanyard, button gap, door, white border, letterboxed

It's not that good with the character, so have to add blunt bangs, and switch bow hairband to hair ribbon and negative hair intakes to get a more accurate hair band. Took me too long to remember to add copyright name, which mostly fixes the hairband but still comes out a little messed up half the time. It likes the notice lines for some reason. To avoid the doorway framing, have to add at least one other thing related to background like window.
>>109973440
engagement baiter, otherwise (NTA) the s-123 at the end of filename is NAI which has stealth metadata
>>
what is a good local model or whatever to create cute chibis that look sketch-ish and not hyper colored and ai slopish
>>
>>109974179
ummm, anima?
>>
Boomer here trying to get started with this stuff
My PC is a dinosaur, can't really rely on it for AI. I want to start making images using Cloud services. Dont mind paying unless its way too expensive (100$+)
What I need is trainable models for consistency (lewd stuff, obviously). Dont mind using already made ones.
Where do I start?
>>
File: ComfyUI_temp_zmebz_00009_.png (2.45 MB, 1024x1536)
2.45 MB PNG
>>109974471
I don't know about renting hardware; if I were in your shoes I would talk to one of the robots about that. As to what model to use, it depends on your preferences; as you're asking here I'm assuming you prefer anime style, in which case Anima is probably your best bet. If you want something more generalist, Krea is generally better, it just doesn't have as much anime knowledge. Both of them can run on pretty mediocre setups so you don't need to rent anything high-end. Any other advice depends on which model you choose.
>>
>>109974694
Cute gen.
>>
>>109974694
Do you guys really just gen stuff like this?
>>
>>109974471
Just play around with grok until you get used to promting first. If you fail at that just look for sloppers on X, select the styles you like, a lot of them take on commisions for NSFW through patreon within reason and some do video too.
>>
>>109974961
What do you mean?
>>
File: AniStudio-05946.png (1.31 MB, 1024x1536)
1.31 MB PNG
>>109974961
what do you gen?
>>
>>109975014
I gen light ecchi and post it on pixiv. I want people to see my gens.
>>
File: AniStudio-03692.png (1.36 MB, 1408x1024)
1.36 MB PNG
>>109975032
That is admirable. Would be nice to see some of your work here too. I just want to make my app a lot easier for people to use all this stuff instead of relying on python for everything
>>
has anyone ever got banned on modal, I wonder what it takes
>>
>>109975081
schizo
>>
File: ComfyUI_temp_afqjl_00111_.png (1.57 MB, 1024x1536)
1.57 MB PNG
>>109974843
Thank

>>109974961
I mean it's my gen so yes I suppose I did do that
>>
File: pine computer.png (3.03 MB, 1856x1536)
3.03 MB PNG
>>109974961
you're not really alive if you don't gen things to memepost
>>
>>109932045
TOTAL YUNYUN CUDDLES AND LOVE!!!
>>
File: CuteCuteCuteVeryCute.webm (1018 KB, 576x736)
1018 KB
1018 KB WEBM
>>109975558
>oh! is Yunyun!
>>
File: CuteCuteCute!.webm (664 KB, 576x736)
664 KB
664 KB WEBM
>>109975693
>>
File: ComfyUI_00075.png (2.77 MB, 1448x1448)
2.77 MB PNG
Trying to make ming-image-0.1 gen a Kuroki Tomoko without the benefit of a LoRA. It seems kinda flux-ish, but not as bad. ai-toolkit now supports ming-image, so I'll be doing a LoRA with my Danbooru 40K Q/E dataset tonight.
>>
>>109975950
cute tomo!
>>
File: 261003144400_output-0001.png (1.68 MB, 1192x1744)
1.68 MB PNG
How do you get jewelry, armor, and other fine details to look better? You can inpaint them but it can get troublesome.
>>
>>109976483
Do them simple first then add detail img2img after upscaling
>>
File: 1791074101610.png (3.41 MB, 1200x2288)
3.41 MB PNG
>>
>>
File: Krea2_turbo_01770_.png (1.64 MB, 1024x1024)
1.64 MB PNG
>>
>>109975736
Living the dream huh
>>
File: YunyunTiddies.webm (821 KB, 576x736)
821 KB
821 KB WEBM
>>109977346
>Yunyun that bikini is a little bit small for your sizes
>>
>>
>>
>>109978143
ぷにぷにぽいーん
>>
i recognise like 80% of the artist/artstyles here; do you have anything new or just copypasting prompts like a fiverr hindu?
>>
>nogen opinion
k
>>
File: ComfyUI_10727_.png (2.12 MB, 1600x1600)
2.12 MB PNG
>>
>>
File: mlg.png (1.66 MB, 1920x1080)
1.66 MB PNG
>>
ming-image-0.1 danbooru 40k q/e LoRA training is underway. Due to the huge text encoder, for the first time, it would not fit in 48GB using bf16, so I had to use the ai-toolkit quantize settting for convrot8. It's rolling along at about 30GB, so training on a headless 5090 might also be possible.
Picrel is one of my reference prompts, it's supposed to be Rin from Yuru Camp wearing a micro-bikini. Rin is well-represented in the dataset, as are micro-bikinis, so I'll be watching to see how close it learns those while hopefully also not fucking up the level of detail in everything else. For a 6B model, ming-image looks damn nice... maybe... just maybe it'll turn out to be trainable and we'll have a better anima.
>>
https://civitai.red/models/2908999/krea2-anime?modelVersionId=3357716
is this any good
>>
Here's baseline ming-image-0.1's idea of Megumin. Yeah, it totally doesn't know her, but again, there's tons of detail in the image, and anatomy is good. I'm using the exported baseline here https://www.runcomfy.com/trainer/ai-toolkit/ming-image-0-1-design-lora-training, LR and weight decay are very low, I'm hoping it behaves like krea and learns instead of doing a z-image and just collapsing to the mean.
>>
File: kreaot2-anime.jpg (548 KB, 2160x1280)
548 KB JPG
>>109979746
good? idk
cute? yes
>>
>>109979840
I meant mostly whether it recognizes most characters
>>
>>109979865
oh. no lol.
>>
>>109979916
aww geez
>>
>>109979865
It takes a huge dataset, at least 300k probably more, to get strong recognition by name, and then you have to flatten things so popular characters don't pull the model in their direction. In my 40k image dataset, name recognition is just a "nice if it works" thing, I'm just trying to add NSFW.
>>
Ruh roh! Not a good sign from the first preview at step 1000. What you should see is a gradual progression from the baseline to something close to an actual captioned image in the dataset, but this... I dunno... I'll let it make a few more and see where it goes, but this looks bad.
BTW it's the Megumin prompt.
>>
>>109980072
>>109980072
>>109980072



[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.