Kaiwen Wang1, Frank Bieder2, Yinzhe Shen1, Carlos Fernandez1, Jan-Hendrik Pauls1, Omer Sahin Tas2
1Karlsruhe Institute of Technology (KIT) 2FZI Research Center for Information Technology
Wavelet Phase Diffusion (ψ-PD) bridges the appearance gap in simulation-to-reality translation while strictly preserving structural and semantic consistency. By operating in the Dual-Tree Complex Wavelet Packet Transform (DT-ℂWPT) domain with Low-Frequency Randomization (LFR), ψ-PD achieves spatially localized phase injection without global spectral artifacts or architectural modifications. This localized formulation enables both global and instance-level sim-to-real translation for images and videos, significantly improving photorealism and downstream utility.
@article{wang2026wavelet,
title={Wavelet Phase Diffusion for Structurally and Semantically Consistent Sim-to-Real Translation},
author={Wang, Kaiwen and Bieder, Frank and Shen, Yinzhe and Fernandez, Carlos and Pauls, Jan-Hendrik and Tas, Omer Sahin},
journal={arXiv preprint arXiv:2607.21628},
year={2026}
}Model weights are released under CC-BY-NC-4.0.

