Diffusers adds LoRA fine-tuning example for NVIDIA's RFD4-Proteina protein design model

RisingSayak · x · 2026-11-20

Hugging Face Diffusers merged a LoRA training example for NVIDIA's RFD4-Proteina protein design model. The recipe trains rank-32/alpha-64 adapters on the denoiser with flow-matching losses over backbone coordinates and local latents instead of pixels, keeping the autoencoder frozen — code closely mirrors standard image/video LoRA training. Validation trained on 255 CPSea complexes for 1,000 steps on two H200s, testing PD-1 cyclic peptide inputs (held out from training) at steps 0/500/1,000 with 16 seeds per chemistry. CIF preparation is cached and unusable structures filtered.

Related event: Diffusers Adds LoRA Fine-Tuning Support for NVIDIA RFD4-Proteina(2 posts)→

Original post →

More from Research

Research channel →