PhysStream: Streaming Physics-Grounded Video Generation with Structured Scene Memory and Fine-Grained Motion Control
PhysStream is built to let users steer generated object motion while a video is still being made.
The paper describes an autoregressive image-to-video model that keeps structured memory from prior frames and accepts sparse velocity-increment controls. Its target setting is multi-object tabletop rigid-body scenes, where the authors say prior methods lack mid-generation interactive control. In reported synthetic tests, it cuts FVMD by 33% and trajectory error by 12% against the strongest baselines. Human evaluators preferred it in more than 85% of in-the-wild comparisons. ArXiv · AI/CL/LG's note
The paper describes an autoregressive image-to-video model that keeps structured memory from prior frames and accepts sparse velocity-increment controls. Its target setting is multi-object tabletop rigid-body scenes, where the authors say prior methods lack mid-generation interactive control. In reported synthetic tests, it cuts FVMD by 33% and trajectory error by 12% against the strongest baselines. Human evaluators preferred it in more than 85% of in-the-wild comparisons. ArXiv · AI/CL/LG's note
score 5