Reproduce “DiffusionART: Efficient, Stable and Scalable Reinforcement of Diffusion Models” 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
Zhenglin Wan, Jiayu Chen, Zhengfeiyang Zhang, Pengfei Zhou, Wangbo Zhao, Fangyuan Kong, Pengfei Wan, Kun Gai and 1 more
DiffusionART improves text-to-video generators using output-level rewards. It pools prompt rewards and outputs into a shared direction, applies it to reverse-denoising states, and restores predictions according to their drift. On HunyuanVideo-13B under equal weighting of visual- and motion-quality rewards, final-50-step mean reward is 13.018, versus 10.105 for DiffusionNFT and 8.510 for Flow-GRPO-Fast; GPU-hour gains apply to this experiment.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.