HomoGen: Enhanced Video Inpainting via Homography Propagation and Diffusion

Input with Mask Stage 2: Content Generation Flow-based Warping Homography Registration without consistency check or hand-crafted filtering operation Figure 1. Two-stage pipelines of HomoGen and the state-of-the-art method. Previous flow-based propagation warps pixels across frames, followed by a video transformer (VT) that generates remaining missing content. In contrast, HomoGen's homography-based propagation warps pixels as priors to control content generation in a video diffusion model (VDM) that filters out artifacts in the priors during denoising.