Research map / Video / Video generation
TAGRPO: Boosting GRPO on Image-to-Video Generation with Direct Trajectory Alignment
Jin Wang, Jianxiang Lu, Guangzheng Xu, Comi Chen, Haoyu Yang, Linqing Wang, Peng Chen, Mingtao Chen, Zhichao Hu, Longhuang Wu, Shuai Shao, Qinglin Lu and 1 more
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 J_align trajectory alignment loss (ablation) | TAGRPO on HunyuanVideo-1.5 · component ablation, Q-Save training reward | Q-Save: 8.48 -> 8.66 | Improved |
| add memory bank (ablation) | TAGRPO on HunyuanVideo-1.5 · component ablation, Q-Save training reward | Q-Save: 8.55 -> 8.66 | Improved |
| update memory bank every step instead of every 2 steps (ablation) | TAGRPO on Wan2.2-TI2V-5B · memory bank update ablation | Q-Save: 8.95 -> 9.00 | Improved |
| use FIFO instead of score-driven update (keep only extreme rewards) (ablation) | TAGRPO on Wan2.2-TI2V-5B · memory bank update ablation | Q-Save: 8.85 -> 9.00 | Improved |
| vary memory bank size N from 4 to 16 (ablation) | TAGRPO on Wan 2.2 · memory bank size | Q-Save | No clear effect |
| vary gamma from 1 to 5 (ablation) | TAGRPO on Wan 2.2 · gamma sensitivity | Q-Save | No clear effect |
| replace DanceGRPO with TAGRPO | HunyuanVideo-1.5 I2V · TAGRPO-Bench (320p) | HPSv3: 1.84 -> 2.41 | Improved |
| apply DanceGRPO to HunyuanVideo-1.5 | HunyuanVideo-1.5 · TAGRPO-Bench (320p) | HPSv3: 2.00 -> 1.84 | Hurt |
| replace DanceGRPO with TAGRPO | Wan 2.2 I2V · TAGRPO-Bench (720p) | HPSv3: 4.40 -> 5.03 | Improved |
| replace DanceGRPO with TAGRPO on text-to-video | Wan2.2-T2V-A14B · T2V generalization, HPSv3 | HPSv3: 5.42 -> 6.27 | Improved |
| replace DanceGRPO with TAGRPO | HunyuanVideo-1.5 I2V · TAGRPO-Bench (320p) | Q-Save: 8.01 -> 8.05 | No clear effect |
Closest papers by meaning
- Neighbor GRPO: Contrastive ODE Policy Optimization Aligns Flow Models (CVPR 2026 · similarity 0.90)
- TEMPFLOW-GRPO: WHEN TIMING MATTERS FOR GRPO IN FLOW MODELS (ICLR 2026 · similarity 0.89)
- Flow-GRPO: Training Flow Matching Models via Online RL (NeurIPS 2025 · similarity 0.88)
- VideoDPO: Omni-Preference Alignment for Video Diffusion Generation (CVPR 2025 · similarity 0.87)
- DeepVideo-R1: Video Reinforcement Fine-Tuning via Difficulty-aware Regressive GRPO (NeurIPS 2025 · similarity 0.86)
- DenseDPO: Fine-Grained Temporal Preference Optimization for Video Diffusion Models (NeurIPS 2025 · similarity 0.86)
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.
