Time-Correlated Video Bridge Matching

Diffusion models perform well in noise-to-data generation but are less effective for data-to-data translation tasks. Bridge Matching (BM) addresses this issue, though its application to time-correlated sequences remains unexplored. We propose Time-Correlated Video Bridge Matching (TCVBM), a framework that extends BM to video by explicitly modeling temporal dependencies within the diffusion bridge. We evaluate TCVBM on frame interpolation, image-to-video generation, and video super-resolution, showing improved performance over classical bridge matching and diffusion-based methods across benchmark datasets and human evaluation.