Reflection Removal Using Regression with Dual-Stream Diffusion Model

Abstract

While diffusion-based methods have advanced Single Image Reflection Removal (SIRR) by leveraging generative priors, their reliance on iterative sampling creates a prohibitive computational bottleneck. To address this, we apply the emerging one-step diffusion paradigm to SIRR, using the diffusion U-Net as a deterministic regressor. By employing a latent space regression scheme based on the v-prediction objective, our model maps reflection-contaminated inputs to clean transmission latents in a single forward pass. To mitigate spatial information loss inherent in VAE compression, we introduce an Adaptive Reflection Filtering (ARF) module within the VAE encoder-decoder skip connections to dynamically suppress artifact propagation while preserving high-frequency details. We further scale this approach to a dual-stream architecture where the Mutual Interaction (MI) module facilitates explicit disentanglement of transmission and reflection layers. Extensive evaluations on SIR2, Real, and Nature datasets demonstrate that our dual-stream variant establishes a new state-of-the-art with 27.34 dB PSNR, outperforming existing methods by 0.65 dB. Additionally, our approach achieves better computational efficiency compared to conditional generation-based methods. Our single-stream and dual-stream models reduce computational costs by 11× and 5×, respectively, compared to iterative diffusion methods.