Wavelet Phase Diffusion improves sim-to-real translation by 5%

A new AI research paper titled "Wavelet Phase Diffusion for Structurally and Semantically Consistent Sim-to-Real Translation" presents a method that bridges the appearance gap between synthetic and real-world images while preserving structure and semantics. Traditional approaches either require expensive control modules (conditioning-based), rely on complex data pipelines (paired-data), or lack learned appearance priors (training-free editing). Even the promising phase-preserving diffusion has limitations due to global spectral coupling in Fourier-domain formulations, leading to ringing and boundary leakage.

Wavelet Phase Diffusion solves this by operating in the Dual-Tree Complex Wavelet Packet Transform domain, where localized wavelet packets allow spatially adaptive phase injection without global interference. Additionally, Low-Frequency Randomization (LFR) replaces the low-frequency packet to remove synthetic illumination priors. The method trains on unpaired open-domain data with negligible inference overhead and enables instance-level translation—translating individual objects or regions to photorealistic appearance independently. On vKITTI→KITTI translation, it outperforms prior methods in realism and semantic consistency. For CARLA video translation, it reduces VLM planner Average Displacement Error (ADE) by 5.4% and Final Displacement Error (FDE) by 5.1%, approaching the realism of paired-data methods.