Structured Diffusion Bridges: Inductive Bias for Denoising Diffusion Bridges

2026-05-03

AI Summary

This work addresses the challenge in modality translation tasks, where the highly under-constrained solution space traditionally necessitates fully paired data, thereby limiting real-world applicability. To overcome this limitation, the authors propose a Structured Diffusion Bridge framework that reframes paired supervision from a strict requirement into an optional alignment constraint. By incorporating structural inductive biases, the method flexibly models cross-modal mappings across unpaired, semi-paired, and fully paired settings. Extensive experiments on both synthetic and real-world benchmarks demonstrate its robustness: even with substantially reduced reliance on paired data, the approach achieves generation quality approaching that of fully supervised methods, significantly enhancing the practicality and generalization capability of modality translation systems.

Abstract

Modality translation is inherently under-constrained, as multiple cross-modal mappings may yield the same marginals. Recent work has shown that diffusion bridges are effective for this task. However, most existing approaches rely on fully paired datasets, thereby imposing a single data-driven constraint. We propose a diffusion-bridge framework that characterizes the space of admissible solutions and restricts it via alignment constraints, treating paired supervision as an optional heuristic rather than a prerequisite. We validate our method on synthetic and real modality translation benchmarks across unpaired, semi-paired, and paired regimes, showing consistent performance across supervision levels. Notably, \textbf{it achieves near fully-paired quality with a substantial relaxation in pairing requirements, and remaining applicable in the unpaired regime}. These results highlight diffusion bridges as a flexible foundation for modality translation beyond fully paired data.

Problem

• modality translation
• under-constrained
• diffusion bridges
• paired data
• cross-modal mapping

Innovation

• structured diffusion bridges
• modality translation
• alignment constraints
• unpaired learning
• inductive bias