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".
EDIT: Why Krea2?
It knows everything. Default quality is off the roof already. (Of course, it's a large 12B commercial model.)
It knows anime/art design, even knows popular characters. (Means it has been trained on a much more creative art dataset. Previous open source models didn't, they mainly for photorealistic)
Official turbo model, 8-step. (FYI: Turbo model is as fast as non-turbo 20-step Anima.)
If you want anime art. 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 (14)
Looks great. Really good work well done 👍. Have a ever considered collaborating with @duongve13112002 to make a solid, stable and high quality Preview 2 checkpoint? As I love preview 1 but the natural language prompt adherence is better especially for more complex scenes and with multiple subjects on P2.
Any examples? I tested p2 and I think prompt adherence is the same as p1, afaik, they were still using tag based captions. And RDBT has better prompt adherence than p2, at cost of built-in styles.
the original p1 does have some stability issues, but it is a "pretrained" model, it should be unbiased and thus unstable. It's ideal for finetuning, but not ideal for users.
I think p2 is a finetuned model, which is difficult to finetune again. Some people said it takes longer to train LoRA on p2, that might be the prove.
@reakaakasky it's been a while since I used it but what is your opinion on NetaYume Lumina aa I think it's great I don't know how it compares to anima tho.
Lumina has way better aesthetics than Cosmos predict 2.
Cosmos predict 2 has better logical understanding.
How are you training these ? LCM ?
just cfg distillation
I see , thank you
Was just curious to try as well.
Tried LCM but it didn't turn out that well , it may need same dataset as original model was trained on.
Works quite well with https://civitai.com/models/2466415/cosmos-predict25-2b-base-distilled-extracted-dmd2-lora at 0.7 strength, 12 steps cfg 1, and https://github.com/pamparamm/ComfyUI-ppm for somewhat working negatives
(silly example prompt: "2girls, kissing (score_9, blushing, :-1.0)" )
very cool
Do you have an fp8 version of anima preview 2?
no, I gave up. it's slower in ComfyUI. fp8 needs torch.compile to inline kernels. Right now torch.compile is unusable, and will be unusable forever if they enforce their dynamic vram mode. On my hardware it's even slower 20% than bf16 +compile.
https://huggingface.co/Bedovyy/Anima-FP8/tree/main
Not mine, not tested, enabled hw fp8, but no calibration metadata.
@reakaakasky Yea, i'm rolling now with silveroxides int8 quant. Slightly faster than fp8 while having basically bf16 quality. 2.34s/it vs. 1.58s/it on my hardware fp16 preview2 vs int8 preview2, no torch compile. Seeds look fairly similar. Btw you were right about klein anime finetune in the making. Apparently chenkin is at it.
. https://huggingface.co/silveroxides/Anima-Quantized/tree/main
@deitychaser I don's see the chenkin klein finetune on their page, or did they just start?
@mc355168 Yea, its still in trainig in testing.
Man all of my lora floading here since your checkpoint is my favorite and use it to generate sample for them lol Still enjoying the checkpoint very much thankfull for it !