PhysStream Ships Autoregressive Physics-Grounded Video Model With Interactive Velocity Control

SIGGRAPH Asia 2026 paper PhysStream introduces an autoregressive image-to-video model that takes sparse user velocity-increment signals to control multi-object rigid-body dynamics interactively, using a structured scene memory of positional and object-tracking maps built online from prior frames. Trained on 100k synthetic indoor videos on a Wan2.2-TI2V-5B backbone, PhysStream cuts FVMD from 1183 to 787 and trajectory ADE from 60.91 to 40.24 px versus RealWonder, wins >85% human preference on physical plausibility, and generalizes to elastic and cloth materials plus 300+ frame horizons.