RDBT | Anima

RDBT [Anima]

Personal finetuned model:

• Comprehensive NL captions from LLM, instead of tags in random order.

• Guidance distilled to further improve quality. Also makes the model 2~4x faster.

• No bias, no default style. Highly creative.

I use it as a starting point to stack more style LoRAs.

See this page for update log. See this page for LoRA version (update more frequently).

Sharing merges using this model is not allowed. Known model thieves: NukeA.I (closed-weight merged model on tensorart),

This model is based on

• AnimaYume (hf link) (civitai link).

• Anima pretrained (hf link)

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.

Steps: 12~24. 12 steps is doable, but quality is not guaranteed as there is no step distillation. It's recommended to add 0.2x turbo lora if you need lower steps (8~12).

Prompt

Specific style is required! This model does not provide a default style. You should always prompt specific style (harder, prompt engineering s*cks). Or use a style LoRA (easier, recommended). Otherwise, you will get random/mixed style. This is a feature, not a bug.

Quality tags:

It's recommended to omit all the quality tags, or just keep the "masterpiece", if you're not confident. Omitting those redundant tokens allows LLM to pay more attention on other words.

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.

Training settings:

~10k images finetuning -> guidance distillation

All captions are NL from Google Gemini.

Optimizer: adamw, constant lr 0.00002.

LoRA rank/alpha 24.

Guidance distillation target CFG 4.

Block 0-2 and adaln linear layers are skipped.