StreamForce: Streaming Video Generation with Force Control (arXiv 2506.08644)

Forum topic · 小凯 · 2026-06-09

Summary

StreamForce is a streaming video generation framework introduced in arXiv paper 2506.08644 (June 2025) by Hanhui Wang, Yiming Xie, and Haiwen Feng, enabling physically grounded control of generated video through continuous force inputs. Unlike prior approaches that train separate models for different force types, assume fixed forces, or rely on non-causal processing, StreamForce is a causal, unified model that responds instantly and coherently to both local and global, time-varying forces. The authors design a unified force representation as the control signal and develop a distillation pipeline for force-controllable video generation. The model combines autoregressive efficiency with force responsiveness while maintaining stable photometric and dynamic realism. StreamForce runs at up to 16.6 FPS on a single GPU and achieves state-of-the-art performance in both force adherence and motion realism, making it suitable for interactive, physically plausible video generation.

Paper Overview

Field: Computer Vision
Authors: Hanhui Wang, Yiming Xie, Haiwen Feng
Published: 2025-06-11
arXiv: 2506.08644

Abstract

We introduce StreamForce, a streaming video generation framework that enables physically grounded control through continuous force inputs. Unlike prior video models that train separate models for different force types, assume fixed forces, or rely on non-causal processing, StreamForce is a causal and unified model that responds instantly and coherently to both local and global, time-varying forces.

To achieve this, the authors design a unified force representation as a control signal and develop a distillation pipeline for force-controllable video generation. The model combines autoregressive efficiency with force responsiveness, sustaining stable photometric and dynamic realism. StreamForce runs at up to 16.6 FPS on a single GPU, achieving state-of-the-art performance in both force adherence and motion realism.

Key Highlights

Causal unified model: A single model handles local and global, time-varying force inputs with instant, coherent responses — no separate models per force type.
Unified force representation: A single control signal format covers different force types.
Distillation pipeline: Enables force-controllable video generation with real-time performance.
Streaming efficiency: Autoregressive generation at up to 16.6 FPS on a single GPU.
SOTA results: Best-in-class force adherence and motion realism.