Discrete Diffusion Bridges for Spatiotemporally Aligned Image Translation and Generation
🌟 Overview
DDB (Discrete Diffusion Bridges) is a framework for spatially and temporally aligned image translation and generation. It introduces:
- A hybrid absorption mechanism that mixes source-image and mask tokens, preserving source structure as spatial anchors.
- An information-guided noise schedule that aligns training corruption with the easy-first decoding process used at inference time.
DDB supports text-guided image editing, structural image translation, and text-to-image generation.