Diffusion Bridges Achieve Near Fully-Paired Quality with Relaxed Pairing Requirements

A novel approach for modality translation utilizing diffusion bridges has been introduced, defining the realm of acceptable solutions and limiting it through alignment constraints. In contrast to many current methods that depend on completely paired datasets, this technique considers paired supervision as non-mandatory. It attains nearly fully-paired quality while significantly easing pairing restrictions and is effective in unpaired scenarios. The framework was validated on both synthetic and real-world benchmarks for modality translation across unpaired, semi-paired, and paired settings, demonstrating reliable performance. This research emphasizes the adaptability of diffusion bridges as a solid basis for modality translation.