RL for Video Generation
RL: Reinforcement Learning
Overview
The field uses reinforcement learning to post-train pretrained diffusion, flow-matching, and autoregressive video generators, aiming to improve outputs beyond what supervised training alone achieves. Common approaches include Group Relative Policy Optimization (GRPO), preference-based optimization, reward-weighted velocity updates, and planner-executor systems that break complex generation tasks into steps. Rewards may score human preferences, text alignment, temporal and geometric consistency, physical plausibility, or whether generated actions succeed in simulated and real robot tasks; evaluation spans benchmarks such as VBench and VideoPhy as well as human studies and manipulation tests. A central challenge is making rewards reliable and informative without encouraging reward hacking, while balancing visual quality with coherent motion, physical realism, and stable, efficient training.
Latest papers
48 papers14 papers in the last four weeks, up 250% on the four weeks before. 0.1% of all new papers.