RadixArk adds LoRA SFT to Miles-diffusion for targeted post-training

Introducing LoRA SFT in Miles-diffusion for fast, targeted post-training of diffusion models.

We trained a rank-64 LoRA for @MiniMax_AI H3 to improve physical realism. Starting from 254 curated training windows, training took under 3 hours on 8 GPUs.

The result is a lightweight adapter that improves targeted capabilities without retraining the full model and can be exported to safetensors and served directly with SGLang.

Follow the guide in the comments to try it yourself ⚡