RVM: Reward-Based Velocity Matching for Diffusion RL Fine-Tuning

GaTech researchers propose RVM, a reward-based velocity matching method that fine-tunes diffusion models with RL directly in velocity space, bypassing likelihood ratios and demonstrating faster, more stable training on SD3.5, Wan and SkyReels.

2026-10-13 ~ 2026-10-14 · 2 related posts

• RVM: reward-based velocity matching fine-tunes diffusion models with RL without likelihood ratios — ArashVahdat · 2026-10-13
• RVM: Likelihood-Free Reward-Based Velocity Matching for Scaling Diffusion RL — gowthami_s · 2026-10-14