Key Info

Hugging Face Diffusers now includes an example for LoRA fine-tuning Nvidia's RFD4-Proteina model for protein design, training rank-32 adapters on the denoiser with flow-matching losses.

Highlights

  • The example trains rank-32 LoRA adapters on the denoiser using flow-matching losses.
  • Losses are computed on backbone coordinates and local latents rather than pixels.
  • Most of the training code closely mirrors standard image and video generation workflows in Diffusers.
  • Details and code are available via the linked example.