High-Fidelity Video Object Insertion with Motion Control
Inserting a reference object into an existing video while preserving its fine appearance details and giving the user precise control over its motion. Built on a text-to-video diffusion model, the approach injects global identity with an ID extractor, controls coarse motion with a box sequence, and adds a pixel warper that warps reference-image pixel details along user-specified key-point trajectories and fuses them into the diffusion U-Net, trained on a mix of videos and static images. Supports downstream tasks like video face swapping, virtual try-on, and multi-region editing without task-specific fine-tuning.
VideoAnydoor: High-fidelity Video Object Insertion with Precise Motion Control Yuanpeng Tu1,2∗ Hao
Inserting a given object into an existing video is hard because the method must both preserve the reference object's fine appearance and produce coherent, controllable motion. VideoAnydoor is a zero-…