RDBT [Anima]

Finetuned + distilled model. Doesn't have a default style. I use it to stack style LoRAs.

See Update Log section for version info. See this page for LoRA version.

All cover images are "raw" output, 1024px, no editing/upscale etc. Metadata included.

Sharing merges using this model is not allowed. It has special trigger words. There is no false positive. Known model thieves: NukeA.I (closed-weight on tensorart)

Usage:

Settings:

CFG scale: 1~4. This model has been guidance distilled. You can disable CFG (CFG 1) and run the model 2x faster. Cover images are without CFG for demonstration.

Prompt

Specific style is required! This model does not provide a default style. You should always prompt specific style. Or use a style LoRA. Otherwise, you will get random/mixed style. This is a feature, not a bug. I use this model as a starting point to stack more style LoRA.

(v0.32+) There are some "roughly classified" trigger words, they are trained so they have effect, but they are not "specific style":

• @anime sketch: Low complexity. Rough outlines.

• @digital anime illustration: Typical "anime". Clear and fine outlines. General complexity.

• @digital art: More complex lighting, textures than typical "anime".

• @cinematic digital art: More lighting, postprocess effects, semi-realistic, etc.

Quality tags:

It's recommended to omit all the quality tags, or just keep the "masterpiece".

Quality tags have been reinforced during distillation. Thus they don't have noticeable effects. Same as negative tags. If you use cfg, there is no need to dump "score_1, blurry, worst quality, jpeg artifacts, extra arms,... x100 words" in your negative prompt. Those things have been distilled out.

Omitting those redundant tokens also allows LLM to better focus its attention on other words.

Training:

Anima pretrained base ckpt -> 10k general image finetuning -> 500 high aesthetic images finetuning -> guidance distillation.

All captions are NL from Google Gemini.

Optimizer: adamw, constant lr 0.00002.

Guidance distillation target CFG 4.

Block 0-2 and adaln linear layers are skipped. Those are much more sensitive. Usually I won't train them for better compatibility (just intuition, no experimental verification).

Update Logs

(5/18/2026): b1 v0.35:

No step distillation. Just guidance distilled.

I'm dropping step distillation. Anima official has their plan to do step distillation (aka, turbo, 4/8-step, or whatever). They have the money and recourse and full dataset. I don't. And my cheap step distillation is kind of sh*tty, tbh.

If you need higher stability or speed, you can stack the extracted cosmos dmd2 lora the anima-turbo, basically can achieve the same thing, probably even better. I prefer 0.2x cosmos dmd2 lora.

(5/12/2026): p3 v0.32.b:

Less step distillation (means higher diversity but less stability). 12 steps is still doable, 24 steps is recommended for complicated prompt.

Styles reinforcement learning. I did this in v0.29, but not in v0.32.

(5/10/2026): p3 v0.32:

No more green-ish, color shifting.

Trigger words have been reclassified to avoid model learning a unified style. See updated "Usage" section.

Old trigger words for backup (v0.29 and before):

• "digital anime illustration": common 2d anime.

• "digital art", 2d art but not anime, mostly digital art.

• "anime sketch": simplified/unfinished anime drawing.

(4/27/2026): p3 v0.29: Distillation algorithm was almost completely rewritten.

Increased diversity. This also improved lighting range, styles and LoRA compatibility.

Better details. This version can squeeze every single pixel out of the VAE.

(4/23/2026) p3 v0.27: Improved stability, details.

(4/18/2026) p3 v0.25: It's based on anima p3.

Previous testing versions, see this page