Research map / Video / Video generation
Flash-GRPO: Efficient Alignment for Video Diffusion via One-Step Policy Optimization
Xiaoxuan He, Siming Fu, Zeyue Xue, Weijie Wang, Ruizhe He, Yuming Li, Dacheng Yin, Shuai Dong, Haoyang Huang, Hongfa Wang, Nan Duan, Bohan Zhuang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add iso-temporal grouping to naive single-step GRPO (ablation) | single-step GRPO on Wan2.1-1.3B · HPSv3 eval reward | eval reward: 4.64 -> 5.31 | Improved |
| add temporal gradient rectification on top of iso-temporal grouping (ablation) | Flash-GRPO · HPSv3 eval reward | eval reward: 5.31 -> 5.42 | Improved |
| add temporal gradient rectification (normalize lambda(t)) (ablation) | Flash-GRPO · training stability | gradient norm spikes | Improved |
| naive single-step training (no ITG, no TGR) vs unaligned base (ablation) | Wan2.1-1.3B · HPSv3 eval reward | eval reward: 4.67 -> 4.64 | No clear effect |
| Flash-GRPO single-step instead of Flow-GRPO-Fast1 sliding window | Wan2.1-1.3B · VBench Imaging Quality | imaging quality: 65.96 -> 68.28 | Improved |
| Flash-GRPO instead of full-trajectory Flow-GRPO | Wan2.1-1.3B · HPSv3 eval reward, first half of schedule | peak eval reward: about 5.1 -> about 5.4 | Improved |
| Flash-GRPO instead of Flow-GRPO-Fast1 | Wan2.1-1.3B · Motion Quality (VideoAlign) | motion score: -0.34 -> -0.28 (approx.) | Improved |
| Flash-GRPO instead of Flow-GRPO | Wan2.1-14B · training reward curve | training speed: 6x acceleration in training cost | Improved |
| Flash-GRPO instead of Flow-GRPO | Wan2.1-1.3B · VBench Imaging Quality | imaging quality: 68.60 vs 68.28 | Mixed |
Closest papers by meaning
- TMPO: Trajectory Matching Policy Optimization for Diverse and Efficient Diffusion Alignment (NeurIPS 2026 · similarity 0.88)
- Co-GRPO: Co-Optimized Group Relative Policy Optimization for Masked Image Generation (NeurIPS 2026 · similarity 0.88)
- Flow-GRPO: Training Flow Matching Models via Online RL (NeurIPS 2025 · similarity 0.88)
- DenseDPO: Fine-Grained Temporal Preference Optimization for Video Diffusion Models (NeurIPS 2025 · similarity 0.87)
- VideoDPO: Omni-Preference Alignment for Video Diffusion Generation (CVPR 2025 · similarity 0.87)
- Inference-Time Text-to-Video Alignment with Diffusion Latent Beam Search (NeurIPS 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
