Paper Overview
- Research Area: cs.CV
- Authors: Shaowei Liu, Xuanchi Ren, Tianchang Shen
- Release Date: 2025-04-09
- arXiv: 2504.06855
- Problem identified: Prior motion-controlled video generation methods cannot cleanly separate object motion from camera viewpoint changes, and they fail to propagate user-driven actions into consistent reactive behavior across the scene.
- Proposed solution: MoRight unifies disentangled motion modeling and causal interaction modeling in a single framework.
- Capabilities:
- Independent control of object trajectories and camera viewpoints
- Causally consistent reactions from non-driven objects in response to user-specified motions
- Experiments: Evaluated on three benchmarks covering video quality, motion controllability, and interaction awareness.
- Results: Achieves state-of-the-art performance across generation quality, motion controllability, and interaction awareness metrics.
- Paper: https://arxiv.org/abs/2504.06855
Abstract
Generating motion-controlled videos requires two capabilities: (1) disentangled motion control, which allows users to independently control object motion and adjust camera viewpoints, and (2) motion causality, which ensures that user-driven actions trigger coherent reactions from other objects in the scene. Existing methods fall short on both fronts.To address these limitations, the authors introduce MoRight, a unified framework built on motion-disentangled modeling. By separating object motion from camera motion and explicitly modeling causal interactions, MoRight gives users fine-grained, independent control over each motion component while preserving realistic downstream reactions.