PhysStream is an autoregressive video generation model that uses structured scene memory and sparse velocity-increment signals to enable fine-grained, physics-aware control during video synthesis. It supports interactive, mid-generation manipulation of multi-object rigid-body scenes without requiring full control sequences upfront.
Why It Matters
This approach enables more natural and precise interactive control in dynamic scene generation, which is critical for applications like simulation, robotics, and virtual environments where physical realism and real-time responsiveness are essential.
Key Findings
PhysStream achieves improved physical consistency by integrating online-derived positional and tracking maps as structured scene memory.
It supports fine-grained motion control using sparse velocity-increment signals that encode physical dynamics, enabling interactive adjustments during generation.
The model reduces motion distribution distance by 33% and trajectory error by 12% compared to state-of-the-art methods on synthetic benchmarks, with strong human preference in real-world evaluations.
Topics
autoregressive
physics-grounded
structured scene memory
motion control
video generation
Thank you to arXiv for use of its open access interoperability. arxiv.org. aibrief.ai is not affiliated with or endorsed by arXiv.