Episode
Flash-GRPO: Efficient Alignment for Video Diffusion via One-Step Policy Optimization
- Podcast
- Daily Paper Cast
- Published
- May 19, 2026
- Duration seconds
- 1264
- Processing state
not_requested- Canonical source
- https://share.transistor.fm/s/48be6e53
Actions
POST https://stenobird.com/v1/public/podcasts/daily-paper-cast-7079649/episodes/flash-grpo-efficient-alignment-for-video-diffusion-via-one-step-policy-optimization/transcription-requests
Idempotently request low-priority transcript generation for this episode.GET https://stenobird.com/podcast/daily-paper-cast-7079649/flash-grpo-efficient-alignment-for-video-diffusion-via-one-step-policy-optimization.md
Read the agent-friendly Markdown representation of this episode resource.
Summary
🤗 Upvotes: 29 | cs.CV Authors: Xiaoxuan He, Siming Fu, Zeyue Xue, Weijie Wang, Ruizhe He, Yuming Li, Dacheng Yin, Shuai Dong, Haoyang Huang, Hongfa Wang, Nan Duan, Bohan Zhuang Title: Flash-GRPO: Efficient Alignment for Video Diffusion via One-Step Policy Optimization Arxiv: http://arxiv.org/abs/2605.15980v1 Abstract: Group Relative Policy Optimization has emerged as essential for aligning video diffusion models with human preferences, but faces a critical computational bottleneck: training a 14B parametered model typically demands hundreds of GPU days per experiment. Existing efficiency methods reduce costs through sliding window subsampling training timesteps, but fundamentally compromise optimization, exhibiting severe instability and failing to reach full trajectory performance. We present Flash-GRPO, a single-step training framework that outperforms full trajectory training in alignment quality under low computational budgets while substantially improving training efficiency. Flash-GRPO addresses two critical challenges: iso-temporal grouping eliminates timestep-confounded variance by enforcing prompt-wise temporal consistency, decoupling policy performance from timestep difficulty; temporal gradient rectification neutralizes the time-dependent scaling factor that causes vastly inconsistent gradient magnitudes across timesteps. Experiments on 1.3B to 14B parameter models validate Flash-GRPO's effectiveness, demonstrating substantial training acceleration with consistent stability and state-of-the-art alignment quality.