PRISM: Latent Composition Consistency for Single-Image Reflection Removal

2026-06-30

AI Summary

Single-image reflection removal is highly ill-posed and suffers from poor generalization due to the nonlinear sRGB color mixing process. This work proposes a novel paradigm that reformulates the problem as a linear separation task in the latent space of a pretrained variational autoencoder (VAE). Leveraging a FLUX backbone, the method learns a velocity field via flow matching to simultaneously recover transmission and reflection layers in a single forward pass. To enable semantic-level disentanglement without ground-truth reflection supervision, it introduces Latent Layer Composition Consistency (LCC) and Layer Contrastive Separation (LCS) loss, integrating latent layer swapping synthesis with cycle-consistency constraints. The approach significantly outperforms state-of-the-art methods across six benchmarks, demonstrating superior performance and strong generalization to real-world scenarios.

Abstract

Single-image reflection removal (SIRR) seeks to recover the transmission layer from a mixture corrupted by reflections -- a severely ill-posed problem. Existing methods operate in pixel space, where the nonlinear sRGB formation model entangles the two layers and limits generalization. We observe that pretrained VAE latent spaces exhibit substantially lower coherence between image layers compared to pixel space, providing a more favorable working space for decomposition. Building on this finding, we propose \textbf{PRISM} (Pretrained-latent Reflection Image Separation Model), which reinterprets SIRR as a latent linear separation problem. Under an approximate additive formulation in latent space, PRISM learns a flow matching velocity field on a pretrained FLUX backbone that recovers both transmission and reflection in a single forward pass. To enforce robust disentanglement, we introduce a Latent Composition Consistency (LCC) strategy that constructs synthetic mixtures by swapping reflection latents across samples and enforces consistent decomposition via a cycle loss. We further propose a Layer Contrastive Separation (LCS) loss that promotes semantic separation between layers through patch-level contrastive learning, without requiring explicit reflection targets. Experiments on six benchmarks demonstrate that PRISM consistently outperforms state-of-the-art methods by significant margins, with strong generalization to in-the-wild images.

Problem

• single-image reflection removal
• transmission layer recovery
• image decomposition
• ill-posed problem

Innovation

• latent space decomposition
• flow matching
• Latent Composition Consistency
• Layer Contrastive Separation
• single-image reflection removal