Wavelet Phase Diffusion for Structurally and Semantically Consistent Sim-to-Real Translation

The paper introduces Wavelet Phase Diffusion for sim-to-real image and video translation. It replaces globally coupled Fourier-domain phase injection with localized phase manipulation in the Dual-Tree Complex Wavelet Packet Transform domain, and uses Low-Frequency Randomization to reduce dependence on synthetic illumination priors. The method trains on unpaired open-domain data, adds negligible inference overhead, and supports instance-level translation. On vKITTI-to-KITTI translation, it reportedly improves realism and semantic consistency while preserving structure. On CARLA video translation, it reduces a VLM planner’s ADE and FDE by 5.4% and 5.1%, respectively.