Latest: (4/14/2026) v0.24f dmd2 b:

Comparing with "v0.24f dmd2": Some adjustments in distillation target. Improved details and stability, images should be sharper. Settings about dmd2 weren't changed, still considerably "lower" than normal 8 steps dmd2 (That's why I called it 16 steps dmd2).

Cover images are using 8 steps for demonstration. I still recommend 12~16 steps for better quality.

This should be my last update before next anima release. I'm quite happy about the quality, and out of juice and idea.

See Update Log section for previous versions.

RDBT [Anima]

Finetuned circlestone-labs/Anima.

Dataset: ~60k images. Zero AI slop. Natural language from Gemini. Includes common enhancements such as eyes, faces, hands, clothes, lightings, backgrounds, etc.

Trained as a LoRA for better training and distribution efficiency. Then cfg/dmd2 distilled for better stability and quality.

No overfitted default style. Still creative, but more stable and aesthetic.

Sharing merges using this model is not allowed. If you think this LoRA is useful, please share the link or the LoRA file.

Usage:

Base ckpt:

The base ckpt that trained the LoRA is pretrained anima base ckpt. Official HF link.

If you just want a "finetuned base ckpt" and forget about LoRA, you can download this ckpt, which has merged the LoRA.

You can also use other further "trained" base ckpts, like AnimaYume, Hikari Anima ... "trained" ckpt has stable (mathematically perfectly optimized) weights and are very close to the pretrained ckpt. No compatibility issues.

Do not use "merged" ckpts. 1) Weights are not optimized. 2) Many of them also merged distillation model without any notice. Their weights are in a mess. If anything happened with a "merged" ckpt, blame the "merged" ckpt, not the LoRA.

Prompt

Prefer natural language prompt. Prompt structure: style, subject, action, background.

You can omit all the quality tags. The quality of training data is higher than "masterpiece". Quality tags don't have noticeable effects.

There are two "rough" trigger words:

• "digital anime illustration": 2d anime.

• "digital art", 2d art but not anime, mostly digital art. (not many samples)

Recommended settings:

• "er_sde", "euler" or "euler_a" sampler.

• dmd2 distilled: 8~16 steps. cfg distilled: 20~30 steps.

• cfg scale 1~3. Prefer cfg 1 (disable cfg, smooth sampling process, 2x faster). Enable cfg (cfg >1) if you need higher prompt adherence (e.g. style is too weak). High cfg is not necessary. Cover images are all using cfg 1.

More effects: In order to make this model stable. I moved some training images to this LoRA. They have very cool effects/features, but too creative/chaotic to be descripted, so they must be trained as a separated LoRA.

FAQ:

cfg distillation and dmd2 :

• cfg distillation: Basic. Has stable sampling process thus has better textures and details. Small stability improvement. E.g.: img.

• dmd2 (Improved Distribution Matching Distillation): The most popular one. cfg + step distillation. Huge stability improvement.

I recommend dmd2 to most users. Which has higher stability and overall quality, also faster.

Use cfg distilled version if you prefer pure aesthetic/creative mode and can handle stability issues.

Update log

[base model version] [finetuned model version] [distillation method] [distillation version]

f = finetuned, d = cfg distilled, dmd2 = distribution matching distillation.

Recommended versions:

• p3 v0.24f dmd2: 16 steps dmd2.

• p3 v0.24fd b: cfg distilled.

• p2 v0.23f dmd2 b: 4 steps dmd2. +200% stability

===============

(4/11/2026) p3 v0.24fd b: Fixed bug in previous 0.24fd.

(4/10/2026) p3 v0.24f dmd2: roughly 12~16 steps distillation. This is intentional. Low steps distillation = ai slop style without complex texture.

(4/8/2026) p3 v0.24fd: Rebased on preview3. Finetuned base model has -40% steps than v0.23. Less overfitted (?).

Update: There is a bug in distillation that causes huge quality downgrade when using without cfg.

(4/4/2026) p2 v0.23f dmd2 b: Different distillation settings. Almost a 4 steps dmd2. Maximum stability.

(4/4/2026) p2 v0.23f dmd2: 8 steps dmd2. First dmd2 anima (?).

(3/28/2026) p2 v0.23fd: Rebased on preview2. Distillation: improved small details and stability (removed a regularization in distillation target and changed to second-order method).

Voting result: v0.20fd won. Thanks for the feedbacks.

(3/24/2026) preview1 v0.20fd b: Distillation: Different settings optimized for anime, high contrast and saturation.

(3/23/2026) p1 v0.20fd: Dataset: More furry. Finetuned base model: from v0.12 + more 100% steps. Distillation: Fixed noisy pixels this time, really.

(3/14/2026) preview v0.19fd b:

Updated dataset. Some private datasets have been dropped. You might notice the style changed.

Fixed high-freq artifacts in v0.12, now you should get a clear image without noisy pixels

b: Testing new distillation settings. Higher contrast. Aligned with common anime art.

(2/19/2026) preview v0.12fd:

Better stability and details, extended dataset.

(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.

(2/12/2026) 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.