RDBT - Anima

RDBT [Anima]

A LoRA for circlestone-labs/Anima.

Tiny scale finetune + CFG distilled.

About CFG distilled:

• You should use CFG scale = 1. It is a special value. It means disabling CFG and neg prompt.

• Because you don't need to run a forward pass for the negative prompt, you can generate 2x faster. (Almost as fast as SDXL)

• No free lunch, same as all distilled models, less diversity. As for anime model, the most noticeable thing is that artist styles are weaker. But output should be more stable.

• But I think the trade-off is acceptable. This is not step distillation, which is very aggressive. CFG distillation is simple.

Versions

• v0.2: 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.

Training settings:

• adamw + lr 0.00005 (seems still high?)

• LoRA rank: 24

• Targeted modules: blocks.[5:27].