v1.0: I think it's time to release v1.0. I've migrated to the krea2 model (never imagined that krea2 would be my next choice...). Fine-tuning the anima model is really fun.
I also uploaded int8 version for ComfyUI hardware int8 (labeled as fp8).
About 8/16 step distillation:
8-step has higher stability. Easy to use and prompt.
16-step (same as previous old versions) has higher diversity and more dynamic details, less sloppy but less stable, more small errors. (I still prefer this one because anime/art model needs a little bit creative and chaotic.)
Comparison images are uploaded as 16-step version showcase images. You can feel the "complexity" of 16-step is noticeable higher than 8-step.
FYI: More distillation = higher stability = lower steps = lower diversity/complexity = more slop-ish. For reference, official anima turbo is 4-step distillation, means highest stability and lowest diversity.
Interesting old version:
b1 v0.39.b: compared to v1.0 16-step, more creative, more dynamic details, but more logical errors. Good for creative, abstract effect, etc.
b1 v0.37.1: 2-stage finetuned, highest aesthetic among all versions, but due to small dataset it has bias. Also has several built-in quality/style trigger words, see update log "b1 v0.32".
P.S: Why Krea2? It knows everything. It knows anime/art design, knows popular characters. It was trained on such dataset, and default quality is off the roof already, all you need is a style/character LoRA. Life is much easier...
RDBT [Anima]
This is a finetuned model with 10k high aesthetic images paired with natural language captions from LLM. Then distilled to further improve quality and stability. Dataset does not contain any shiny plastic glossy AI image.
It's not overfitted and doesn't have a default style. I use it as a clean starting point to stack more style LoRAs. I can stack whatever I want and get exactly what I stacked.
See this page for update log.
For advanced users: The RDBT model is trained as LoRA natively. See this page for original LoRA.
This model is based on:
prefix with ym: AnimaYume (hf link) (civitai link).
prefix with b,p: Anima pretrained (hf link)
Sharing merges using this model is not allowed. This "restriction" won't affect anyone. It's only aimed at those who steal others' models to sell.
If someone is selling this model as their own, I'm happy to list them here so everyone knows.
Known model thieves: NukeA.I (selling this model behind paywall on tensorart).
I wrote a story about it. Also contains a guide for trainers about "how to bake special trigger word into your model".
Usage:
Settings:
CFG scale: 1~3. This model has been distilled. You can disable CFG (CFG 1) and run the model 2x faster. Cover images are without CFG for demonstration. "RenormCFG" node is recommended
Steps: 16+
Prompt:
Always specify style, or use a style LoRA. Otherwise, you will get random/mixed style. This is a feature, not a bug. This model does not provide overfitted default style.
Quality tags:
It's recommended to omit ALL quality tags. The fine-tuning dataset has higher quality than "masterpiece". Thus they don't have noticeable effects. Omitting those redundant tokens allows LLM to pay more attention on other words
Training settings
All captions are NL from Google Gemini.
Optimizer: adamw, constant lr 0.00002, weight decay 0.1, batch size 16.
LoRA rank/alpha 24.
Timesteps shift 3.
Block 0-2 and adaln linear layers are skipped.
Description
FAQ
Comments (32)
im using anima model with 4-7 cfg sometime it can get good prompt adherence , does that doing the trick with this ckpt with 1-3 cfg ?
Just reuploaded the same ckpt again. For those who found and downloaded previous ckpt through notification system. You don't have to redownload. Cvitiai model api is still f***ked, previous assigned model ID was desynchronized and didn't get indexed, it disappeared across the site, didn't even show on my own profile page.
Would it be possible for you to adapt the NAG comfynode to support Anima? I remember this was done by a third party for ZIT because the official creator do not update model support of NAG (normalized attention guidance). NAG is allowing us to have negative conditioning even with turbo distilled models at CFG=1, while not really slowing down the model. Imo its even more effective for negative guidance than CFG in normal undistilled models.
In any case would be a great addition to the workflow.
https://github.com/BigStationW/ComfyUI-NAG this the most advanced fork of nag but still does not support anima :(
@brahianvalles It got now anima support!
0.25 version is an actual "RIP illustrious/noobai" the model has gotten too good.
Holy shit, man, you're a monster. Do you even sleep? I can't keep up with testing your models at the rate you release them!
this is actual black magic. how the fuck does this work? how the fuck CAN this work?
besides my bewildered amazement that we got a distilled finetuned model this soon for anima,
it definitely does not handle loras + text very well, text without loras mostly generates just fine like other models but introducing a lora totally destroys its coherency. otherwise, yeah this is pretty amazing it's up there with wai and animaika from what i've tested so far. and doing all that in only 12 steps ,very impressive.
Anima getting too good now
Can you run for president?
Nice update (v0.25), i tried the original one, Yume and WAI, but yours is better just by the low step requirement and i tried some style LoRas and they work well.
The ultimate Anima checkpoint so far. Have you ever thought of integrating the Cosmos-Predict2.5-2B base distilled extracted DMD2 LoRA into your great distillation work.
No. I prefer a clean model status, trained (optimized), no merge conflicts, for better stability.
Adding more LoRAs, that's up to you guys.
@reakaakasky actually I did notice that as I have tried using two DMD2 LoRA two your one and the cosmos one. And I did notice while it is good, especially regarding adding a bit more contrast. It also gives me a little bit more inconsistent results. So yeah I do agree with you on that.
Also the official anima team has released a turbo LORA however, it's not as stable as your one. I've noticed it does allow a little bit more diversity, but regarding stability and actually prompt adherence it's not as good as yours. Just wanted to say.
nice model, thanks
"Sharing merges using this model is not allowed"
But why?
You're not allowed to merge this checkpoint or LORAs with other models and share them I think. That's what it means as far as I'm aware.
@AnimaXx Yeah that's why i'm asking why author decided to not allow merges.
https://civitai.red/articles/26855/dev-story-stabilizer-and-rdbt
might be the answer
@ikekph5 Yeah that's explains, thanks.
For anyone unaware, this model can do photorealism, but it's not particularly great, since it isn't the focus. A lora could maybe make it work.
I keep getting photorealism when I don't want it 😭
Yeah, anima is still very bad with photorealism🫤
Something went downhill with V0.25, it is worse than V0.24
1. Listen to prompt less
2. Adds elements to image that were not in prompt
3. Is unstable
4. At times throws plasticy 3d/2.5d looks when it is clearly against prompt and art style
5. Consistently worse at generating texts in ratio 7:3 (+/-)
6. At times fails at anatomy (melted hand, melted feet)
here are examples/comparisons:
https://civitai.red/posts/28090228
https://civitai.red/posts/28093793
+ Can generate realistic looking people/photo-realism meaning some undesirable material got in and is littering the model. Anima was supposed to be 2D model, please don't do out of it 2.5D or 3D model, model for everything is model for nothing.
Thank you for your work, you made Anima usable model when it can work on CFG1, will cross my fingers and wait for V0.26 (perhaps p4 variant) Keep up the good work ❤️🙏 Muchas gracias
I'm writing an article about dmd2 anime model. It's really a shi*hole. I know lots of people wondering why I keep changing distillation settings drastically, wtf is this guy doing.
Short answer, I was trying to find a solution, but after 1 mouth, I now believe there is no such solution.
How to improve contrast in Forge NEO without increasing CFG above 1
I don't use Forge but try to look for cfg loras, they do exist, it might help x) I have one called "CFG Scale Boost"