Trained with 1024 and 512 simultaneously.
Fully finetuned from Anima-Preview-3.
No need for trigger word.
No separate face folder.
Some tags have been protected.
The original Qwen-Image VAE has been used.
Edit: Trained with only tags, so you might face bad performance when using natural language.
Edit1: Used 2.2e-6 LR with Silvelter's Lion8bitSR optimizer. Anima is chicken, so the model got burnt. The same LR works great on NoobAI XL vpred 1.0, Mugen and SIH.
- Downloads last month
- 123
Model tree for StarkEater/NyteTydeAnima12052026
Base model
nvidia/Cosmos-Predict2-2B-Text2Image Finetuned
circlestone-labs/Anima