jingw193/TempAct

TempAct is a reinforcement learning framework for autoregressive video generation, aiming to improve the rationality of action timing in videos. It jointly optimizes the LLM planner and diffusion executor to ensure that the generated video strictly follows the time sequence in the prompt words (such as pouring the sauce first and then adding the cheese), solving the problem of timing confusion in the traditional method, and is suitable for video generation scenarios that require high-precision event progress control.