Bidirectional Diffusion Bridges Enable Flexible Multimodality Translation

Multimodality translation, such as generating images from text, is a fundamental task in generative AI. Current methods often follow generative paths that don't directly represent the source modality, limiting sampling flexibility, and are typically unidirectional, preventing inverse operations like image-to-text translation. A new study introduces BIT: Bidirectional Image-Text Diffusion Bridges, which addresses these limitations. BIT initiates generation directly from text and interpolates into images, creating a source-aware generative path. This design enables more diverse and flexible sampling algorithms. Furthermore, BIT establishes an endpoint-conditioned process that can be traversed in reverse, from image to text, thereby providing a unified, bidirectional generative framework. Derived from stochastic calculus, BIT offers SDE forms suitable for simulation and tractable loss functions that scale to high dimensions. Experimental results show BIT's competitive performance against denoising-diffusion and deterministic-flow baselines, and superior results in several vision-language and natural-science evaluations.