Reflection Separation from a Single Image
via Joint Latent Diffusion

Single-image reflection separation remains challenging due to its ill-posed nature, especially under extreme conditions with strong or subtle reflections. Existing methods often struggle to recover both layers in glare or weak-reflection scenarios because of insufficient information.

This paper presents the first diffusion model explicitly fine-tuned for this task, leveraging generative diffusion priors for robust separation. Our method simultaneously generates transmission and reflection layers through a unified diffusion model, incorporating a novel cross-layer self-attention mechanism for better feature disentanglement. We further introduce a disjoint sampling strategy to iteratively reduce interference between the layers during diffusion and a latent optimization step with a learned composition function for improved results in complex real-world scenarios.

Extensive experiments show our approach achieves superior separation performance on multiple real-world benchmarks and surpasses state-of-the-art methods in both quantitative metrics and perceptual quality.

Given an input image containing reflections, we encode it into a latent representation and leverage a fine-tuned diffusion model guided by distinct prompts ("Transmission" and "Reflection") to jointly separate both layers. Our framework consists of the following key components:

• Cross-Layer Self-Attention — Queries from each layer attend to keys from both layers, enabling effective information exchange and clearer feature disentanglement.
• Disjoint Sampling — Predicted noise differences between layers serve as mutual negative guidance, iteratively pushing transmission and reflection trajectories apart.
• Fidelity-Guided Feature Modulation (FGFM) — Multi-scale features from the original mixture image counteract color drifts introduced by disjoint sampling.
• Latent Optimization — A learned composition function with test-time optimization refines latent representations, improving separation fidelity while reducing computational cost.

Our method achieves state-of-the-art performance on multiple real-world benchmarks, including Real20, Nature, and SIR² datasets. Key highlights:

• Effectively handles strong reflections by hallucinating missing details in the transmission layer.
• Accurately extracts subtle and weak reflections that previous methods fail to capture.
• Achieves superior results in both quantitative metrics (PSNR, SSIM, LPIPS, DISTS) and perceptual quality.
• First method to demonstrate strong reflection layer recovery, not just transmission layer improvement.