2026-10-11 · America/Los_Angeles · 论文 · #5

VideoAWR: Reinforcing Video Flow Models with Decoupled-Timestep Advantage-Weighted Rectification

Regression-based RL for flow models mandates sampling \emph{without} classifier-free guidance (CFG) to align the rectification target with the reference marginal velocity. Extending Advantage-Weighted Rectification (AWR), which dynamically scales this regression target based on reward, to video generation encounters a cold-start pathology: unguided rollouts collapse into low-fidelity regions where intra-group reward gaps remain wide but no longer track perceptual quality, rendering the optimization target arbitrary. Conversely, utilizing guided rollouts to restore fidelity, while keeping the…

热度 17.7 / 100；排名与评分保留该期记录。