Pinned note:
(3/8/2026) preview v0.11fd 512px:
Prove of concept version for v0.12. Same dataset and settings as v0.12, except it was trained with 512px res.
Released by request, as it might be very useful. Running model in 512px and cfg1 is extremely fast (x10 faster, e.g. 30s -> 3s). If you don't have a beefy GPU. You can use this version to test your ideas/prompts in few seconds.
(2/17/2026): This finetune series probably will not be updated.
Anima is a wonderful model. but it has a very restrictive license.
I'm fine with dual licenses (non-commercial + commercial). We all know that training a model needs lots of $. Commercial license is necessary. Commercial means $, $ means better model.
I didn't expect that they keep the right to "sell" your "non-commercial Derivatives". You don't even have the right to make your "non-commercial Derivatives" non-commercial (copy-left). Because they keep the right to apply their commercial license to your "non-commercial Derivatives".
Unfortunately, this is too restrictive for my personal situation. So, this model will not be further finetuned.
Good things:
v0.12 has met 90% of my expectations. (10% is the color issue, and I don't think this can be fixed, because this might be a problem of the base Cosmos model, see FAQ below)
Many models are coming up. It's still too early to say who is the best. E.g Chroma2. Which should be Apache 2.0. And is based on Flux Klein 4B. Much better than Nvidia Cosmos Predict2.
I will also try to finetune Flux Klein 4B for better anime style transfer. The idea is, you provide 2~4 style ref images, and generate a new image with the style. No more style lora. I don't know if it is going to work, nor how long it is going to take. Stay tuned.
RDBT [Anima]
Finetuned circlestone-labs/Anima. Experimental, but works.
Trained as LoRa, for better training and distribution efficiency.
CFG distilled to further improve quality and stability. Kind of similar to Flux Krea. Expect this was homemade.
Dataset:
Handpicked, only top quality masterpieces. Contains common enhancement such as hands, clothes, lightings backgrounds, etc.
No glossy AI slops in dataset. Glossy Al images are polluting the world, but not on my watch.
Natural language captions from Gemini.
Usage:
You must load the LoRA with strength 1 on it's base model.
If you don't know what this means, or which one is the right "base model", you can download and use this full fp16 model, which has merged this LoRA: https://civarchive.com/models/2356447.
Prefer natural language prompt. Prompt structure: style, subject, action, background.
Omit all the quality tags. You don't need those tags. The average output quality of this model should be higher than so-called "masterpiece".
Use CFG scale 1.
Recommended settings:
"Euler a" sampler.
20~30 steps.
"Adjust Contrast" node to boost contrast.
All cover images are directly from the model, at 1024px, no upscale, no face/hands inpainting fixes. Workflow included.
Restrictions:
Making merges using this model is not allowed. FYI, this model was trained with latent watermark. And an OC from my friend with trigger words.
FAQ
Why v0.12 has low contrast?
v0.12 adjusted distillation method. It aligned the sampling process with the "sweet spot/path" of the model. At the "sweet path", the model knows what to generate and how to generate it better (more stable, better prompt adherence, details).
Why this "sweet path" has low contrast? The base model of Anima, aka the Nvidia Cosmos Predict2, is a model for industrial robotics. It's not a model for aesthetic.
Also, based on my testing, boosting contrast in sampling process is a bad idea. Yes, the model is able to generate high contrast image, but "able" != "good at". The best way is adjusting the image contrast/saturation afterward.
What's the art style?
This is a general finetuned model. The dataset is very diverse. It doesn't have a strong default style and is very creative. You should always specify styles you want in you prompt.
Why LoRA?
I don't have millions of images.
I can save VRAM and you can save 98% storage.
Why CFG distilled?
TLDR: Because distillation can improve quality.
Distillation can "purify" the model and solve many problems. That's why distilled models have become the mainstream. We have seen that people like z-image-turbo, flux klein more than their base versions. Not just because they are faster, but mainly because the image quality is better.
I'm not a gigachad who can do step distillation. So this is only CFG distilled. Which means, you still need high steps.
For example, this is what a 30 steps sampling process looks like. You can find more examples and workflow in cover images.
Up: RDBT v0.12.
Bottom: anima preview CFG 4.
Update log
f = finetuned, d = cfg distilled, preview = base model is anima preview
Recommended:
(2/19/2026) preview v0.12fd:
Better stability and details, extended dataset.
(3/8/2026) preview v0.11fd 512px:
POC version for v0.12. Released by request. Same dataset and settings as v0.12, except it was trained with 512px res.
Running model in 512px and cfg1 is extremely fast (x10 faster, e.g. 30s -> 3s). If you don't have a beefy GPU. You can use this version to test your ideas/prompts in few seconds.
Old:
(2/12/2026) preview v0.6d:
CFG distilled only. No finetuning. Cover images are using Animeyume v0.1.
(2/3/2026) preview v0.2fd:
Speedrun attempt, mainly for testing the training script. Limited training dataset. Only covered "1 person" images plus a little bit of "furry". But it works, and way better than what I expected.