Answers are grounded in this paper and cite their sources.
Video object insertion (VOI) aims to seamlessly composite a user-specified object into an existing video while preserving scene consistency and generating realistic object motion that interacts naturally with the surrounding environment. Despite recent progress, achieving high-quality VOI remains challenging. Most existing approaches perform insertion purely in the RGB domain without explicitly modeling scene geometry, which often leads to depth ambiguity, incorrect occlusion reasoning, and unrealistic object–scene interactions. Moreover, current frameworks provide only limited control over object motion and cannot model movement along the depth axis. In this paper, we propose InDepthVOI, a depth-aware video object insertion framework that explicitly incorporates geometric reasoning into the generation process. Our method employs a dual-branch diffusion transformer that jointly models RGB appearance and depth representations. We further introduce a depth trajectory map, a structured conditioning signal that encodes the spatio-temporal trajectory of inserted objects across both the image plane and depth dimensions, enabling depth-consistent motion control. To train InDepthVOI, we construct a new dataset comprising carefully curated background–composite video pairs and structured depth trajectory annotations. Extensive experiments demonstrate that InDepthVOI significantly outperforms existing methods in both visual realism and controllability, establishing a new state of the art for VOI.
No takes yet. Share an insight, caveat, or question.
Han et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: