RDBT | Anima

RDBT [Anima]

This is a personal finetuned + distilled model.

Dataset contains 10k images paired with natural language captions from LLM. All images are handpicked, finished, aesthetic and have accurate anatomy. Dataset does not contain any shiny plastic glossy AI image (aka. not contaminated by glossy AI style and fixed AI faces).

Then distilled to further improve quality and stability.

Compared with the pretrained anima model, this model has more stable composition, fewer logical errors, better lighting, background, detail. E.g.:

• 1, 2 stable composition, better and logical background, no/less artifacts

• 3 no bias, doesn't change/pollute styles, faces.

Because the dataset is big enough. Model was not overfitted and does not have high bias. It can generalize to almost any kind of art style. I use it as a clean starting point for more style LoRAs.

See this page for update log and version info.

For advanced users: The RDBT model is trained as LoRA natively. See this page for original LoRA.

(New: v1.0) For ComfyUI users: Int8convrot version is recommended.

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:

Recipe: This model + style LoRA you want, normally you can go LoRA strength 1 without issues (unless the style LoRA is super overfitted)

Settings:

• CFG: 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 highly recommended if CFG is enabled (CFG > 1).

• Steps: 16+ (some versions can be 8)

• Sampler: euler (er_sde usually is not recommended for distilled model)

Prompt:

Always specify style in prompt. Otherwise, you will get random/mixed style. This is a feature, not a bug. This model does NOT have overfitted default style (which ignores prompt and is always active).

Quality tags:

Omit ALL quality tags. You don't need those. The fine-tuning dataset has higher quality than "masterpiece". Thus quality tags don't have effects. Omitting those redundant tokens allows LLM to pay more attention on other words.

Misc.

Why does this model exist:

I need a model to assist in production level drawing. This model should have top tier quality, and must be able to reproduce the style I want/trained as accurately as possible.

However, existing models:

1. The quality of pretrained models are not good enough.

2. Fine-tuned models suffer from severe style bias because of small dataset, also contaminated by AI styles and faces.

So my friend and I created this dataset and model. We use this model as a clean starting point for style LoRAs. We can stack whatever we want and get exactly we stacked.

Base model:

• prefix with ym: AnimaYume (hf link) (civitai link).

• prefix with b,p: Anima pretrained (hf link)