RDBT [Anima]

(2/17/2026) Latest: 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".

Personal opinion, that's a little bit greedy. Unfortunately, too restrictive for my personal situation.

So, this model will not be further finetuned.

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 klein 4b. Much better than cosmos pt2.

https://huggingface.co/lodestones/Chroma2-Kaleidoscope

Back to topic, v0.10fd update note:

Tldr:
Much better (?) stability and details (?). This is what sampling
process looks like. Up: rdbt, down: cfg 4. You can find more examples
and workflow in cover images.

Finetuned circlestone-labs/Anima. Experimental, but works

Dataset
contains natural language captions from Gemini. But still contains
danbooru tags. Every image in dataset is handpicked by me. Contains
common enhancement such as clothes, hands, backgrounds.

You must specify styles
in your prompt. Dataset is not small and is very diverse. It won't give
you a stable default style. If you don't specify style, the model will
just give you a random/mixed one. This is intentional.

Always use strength 1. Unless you know what you are doing.

Models are CFG distilled:

• Prefer Euler a sampler.
• Use CFG scale 1 to gen 2x faster.
• Use CFG scale 1~2 to get probably better image.
• Model bias might be amplified. Default style that do not need trigger words (it is bias) might be stronger. E.g. Styles from style LoRA. Styles that need trigger (not bias) might be weaker. E.g. base model built-in styles.

Why LoRA?

• I only have ~20k images. A LoRA is enough.
• I can save VRAM when training and you can save 98% storage and data usage when downloading.

Versions

• f = finetuned
• d = cfg distilled.

Based on anima preview:

(2/12/2026) v0.6d: CFG distilled only. No finetuning. Cover images are using Animeyume v0.1.

(2/3/2026) v0.2fd: finetuning
+ cfg distillation. 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.