Reproduce “TAGRPO: Boosting GRPO on Image-to-Video Generation with Direct Trajectory Alignment” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jin Wang, Jianxiang Lu, Guangzheng Xu, Comi Chen, Haoyu Yang, Linqing Wang, Peng Chen, Mingtao Chen and 5 more
A framework for image-to-video (I2V) generation, TAGRPO, enhances video quality by directly aligning denoising trajectories with high-reward samples and utilizing a memory bank for efficient optimization. The method improved HPSv3 scores on the Wan 2.2 model from 4.34 to 5.03 at 720p resolution and achieved higher generation diversity compared to baseline models.
Key insights
- Directly aligning denoising trajectories, rather than just modulating individual sample probabilities, provides more precise and effective guidance for optimizing I2V models.
- Leveraging the shared structural information within groups of videos generated from the same initial noise offers rich and discriminative signals for post-training refinement.
- A memory bank can efficiently provide diverse off-policy samples, crucial for stable preference-based reinforcement learning in generative models, without incurring high per-step rollout costs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.