TriMotion: Modality-Agnostic Camera Control for Video Generation

TriMotion is a newly proposed modality-agnostic framework for camera-controlled video generation that integrates video, pose, and text inputs into a unified motion embedding space. It utilizes a Motion Triplet Dataset for synchronized supervision and introduces a latent motion consistency objective to ensure generated videos adhere to target camera trajectories without pixel-space decoding. This approach enhances flexibility in applications such as sequential motion composition and cross-modal motion interpolation, making it a significant advancement for practitioners in generative systems.