Title: Decoupling Reflections from Single Images Using a Combined Latent Diffusion Framework

Abstract: Extracting distinct reflection layers from a solitary image proves particularly difficult in challenging environments characterized by strong glare or faint reflections. Current techniques frequently fail to accurately reconstruct both components in these scenarios due to a lack of sufficient data. To address this, we propose a specialized diffusion model, fine-tuned specifically for reflection separation, which utilizes generative diffusion priors to ensure robust performance. By employing a unified diffusion architecture, our approach simultaneously produces the transmission and reflection layers. A key innovation is the integration of a novel cross-layer self-attention mechanism designed to enhance feature disentanglement. Additionally, we implement a disjoint sampling strategy that iteratively minimizes cross-layer interference throughout the diffusion process, alongside a latent optimization phase driven by a learned composition function to boost outcomes in complex, real-world contexts. Comprehensive experiments indicate that our method outperforms existing state-of-the-art techniques across various real-world benchmark datasets.

Project page: https://brian90709.github.io/diff-reflection-separation/

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC