Reproduce “Adaptive Domain Shift in Diffusion Models for Cross-Modality Image Translation” on OpenResearch

About this paper

Zihao Wang, Yuzhou Chen, Shaogang Ren

Researchers from Laplace Lab and the University of Tennessee developed Cross-Domain Translation SDE (CDTSDE), a diffusion framework that embeds dynamic, spatially varying domain-shift guidance directly into the generative process for cross-modal image translation. The method consistently improved image fidelity and semantic consistency while achieving up to 2x faster inference across medical, remote sensing, and industrial imaging tasks.

Key insights

• A dynamic, pixel-wise domain transfer schedule can achieve strictly lower path energy and more efficient transitions compared to global linear schedules under heterogeneous local geometry conditions.
• Embedding geometry-aware domain-shift dynamics directly into the reverse-time SDE drift enables on-manifold updates even with larger step sizes, reducing the "correction burden" on the diffusion model.
• This approach reframes the score model's primary role from global alignment to local residual correction, leading to more stable updates and improved fidelity in cross-modal translation.

Summary by alphaXiv.