RetinexDualV2: Physically-Grounded Dual Retinex for Generalized UHD Image Restoration

Abstract

We propose RetinexDualV2, a unified, physically grounded dual-branch framework for diverse Ultra-High-Definition (UHD) image restoration. Unlike generic models, our method employs a Task-Specific Physical Grounding Module (TS-PGM) to extract degradation-aware priors (e.g., rain masks and dark channels). These explicitly guide a Retinex decomposition network via a novel Physical-conditioned Multi-head Self-Attention (PC-MSA) mechanism, enabling robust reflection and illumination correction. This physical conditioning allows a single architecture to handle various complex degradations seamlessly, without task-specific structural modifications. RetinexDualV2 demonstrates exceptional generalizability, securing 4\textsuperscript{th} place in the NTIRE 2026 Day and Night Raindrop Removal Challenge and 5\textsuperscript{th} place in the Joint Noise Low-light Enhancement (JNLLIE) Challenge. Extensive experiments confirm the state-of-the-art performance and efficiency of our physically motivated approach.

Figures (6)

• Figure 1: Visual results of our proposed method on multiple image restoration tasks for NTIRE 2026 test sets, demonstrating its generalizability across low-light enhancement, dehazing, deraining, and shadow removal.
• Figure 2: Inputs and corresponding priors for each task.
• Figure 3: RetinexDualV2 Overview. Based on retinex theory, it decomposes the UHD image into $R_{eff}$ and $L_{eff}$ and operates on them using 2 sub-networks: Physical-Grounded Scale Attentive MaMBA (PG-SAMBA), and Physical-Grounded Frequency Illumination Adaptor (PG-FIA), respectively. Both branches recieves guidance $p_m$ from the Task-Specific Physical Grounding Module (TS-PGM). Noting that each output from a different level has a convolution layer before it for processing, which wasn't illustrated for simplicity.
• Figure 4: Visual comparison between different architectures on UHD-LL Li2023ICLR.
• Figure 5: Visual comparison between different architectures on 4K-Rain13K dataset chen2024towards.