RVM: reward-based velocity matching fine-tunes diffusion models with RL without likelihood ratios
ArashVahdat · x · 2026-10-13
Researchers propose RVM (Reward-based Velocity Matching), a recipe for RL fine-tuning of diffusion models that applies rewards directly in velocity/score space, skipping likelihood ratios entirely. It is fast, scalable and stable, cheap for video fine-tuning, unifies recent objectives, and fixes the static-video problem via dynamic motion rewards.
More from Multimodal
- 30 browser games that never existed: AI-generated videos rebuilt into playable WebGL games — PurzBeats · 2026-10-13
- Runway Tops $100M Q3 ARR as Co-CEO Co-Writes Essay on World Models — tlakomy · 2026-10-13
- 30 playable WebGL games built by Claude from AI-generated gameplay videos — OdinLovis · 2026-10-13
- Dimillian tests Google Astra: one prompt, exactly the result he wanted — Dimillian · 2026-10-13
- AI-made Christmas movie cost $1.5M instead of $80M, director tells Variety — ThomasSimonini · 2026-10-13
- Mage-Flow W4A8 + ARA release: BF16-like quality at 8.9 GB VRAM — ThaJedi · 2026-10-13