Streaming Video Generation with Streaming Force Control

Streaming Video Generation with Streaming Force Control, by Hanhui Wang, Yiming Xie, Haiwen Feng, Zhaoyang Lv, Shenlong Wang, and Huaizu Jiang, introduces StreamForce, a causal video generation system built to respond to continuous physical force inputs in real time rather than offline trajectory prompts. The core idea is a unified force representation that encodes both global forces like wind and local contact forces like pushes in a single pixel-aligned masked tensor, paired with a force-aware distillation pipeline that transfers force-motion behavior from a bidirectional Wan2.2 5B TI2V teacher into an autoregressive student. Training uses 30K synthetic force-conditioned videos, another 30K Blender-rendered force-changing clips, and about 90K diverse Pexels image-force pairs. On a single H200 GPU, the final model streams at 16.6 FPS at 832×480 with 0.6-second latency. In user studies on 40 test cases, StreamForce outperforms baselines on force-changing global and local control, reaching 86.5% force adherence for global changes and 80.4% for local changes, while its Physics-IQ score rises to 46.31 on local-force evaluation and 40.99 on global-force evaluation. The paper’s novelty is not just controllability, but causal responsiveness: users can change force magnitude or direction mid-generation and the model adapts without breaking photometric stability, while also showing emergent intuitive physics such as mass- and friction-dependent motion.