SLER-IR: Spherical Layer-wise Expert Routing for All-in-One Image Restoration

TL;DR

SLER-IR, a spherical layer-wise expert routing framework that dynamically activates specialized experts across network layers is proposed, and a Spherical Uniform Degradation Embedding with contrastive learning is introduced, which maps degradation representations onto a hypersphere to eliminate geometry bias in linear embedding spaces.

Abstract

Image restoration under diverse degradations remains challenging for unified all-in-one frameworks due to feature interference and insufficient expert specialization. We propose SLER-IR, a spherical layer-wise expert routing framework that dynamically activates specialized experts across network layers. To ensure reliable routing, we introduce a Spherical Uniform Degradation Embedding with contrastive learning, which maps degradation representations onto a hypersphere to eliminate geometry bias in linear embedding spaces. In addition, a Global-Local Granularity Fusion (GLGF) module integrates global semantics and local degradation cues to address spatially non-uniform degradations and the train-test granularity gap. Experiments on three-task and five-task benchmarks demonstrate that SLER-IR achieves consistent improvements over state-of-the-art methods in both PSNR and SSIM. Code and models will be publicly released.

Figures (7)

• Figure 1: Illustration of representative all-in-one image restoration frameworks. (a) Prompt-based methods li2022allpotlapalli2023promptirconde2024instructirluo2023controllingtian2024instruct adapt a unified parameter space via feature modulation. (b) Expert-based methods luo2023wmai2024lorazamfir2024efficientyang2024allyu2024multili2020allzamfir2025complexity activate task-specific experts. (c) Our SLER-IR introduces spherical expert routing and global–local granularity fusion for robust restoration.
• Figure 2: Quantitative comparisons (PSNR$^{1\text{-}5}$/SSIM$^{1\text{-}5}$) of four models (AirNet li2022all, Restormer zamir2022restormer, MoCE-IR zamfir2025complexity, and SLER-IR (ours)) across five image restoration tasks (Derain, Dehaze, Deblur, Low-light, Denoise).
• Figure 3: Overview of the proposed SLER-IR framework. Given an input LQ image, the degradation extractor produces a raw routing vector, which is projected onto the unit hypersphere for cosine-similarity gating, enabling layer-wise expert selection under Stage 1 (Probabilistic Routing) and Stage 2 (Deterministic Routing). In parallel, a Global–Local Map Construction (GLMC) module derives a content semantic patch (CSP) map and a degradation severity patch (DSP) map, which are fused through Content-Guided Degradation Fusion (CGDF) to guide restoration.
• Figure 4: Visualization of dynamic routing trajectories across different degradations. Stage 1 learns degradation-aware routing via probabilistic expert selection, while Stage 2 forms specialized expert paths through deterministic routing across layers for different degradations.
• Figure 5: Comparison of existing degradation routing methods and our hyperspherical degradation representation-based branch selection strategy. (a) Traditional expert-based categorization and assignment neglect inter-degradation similarities. (b) Linear-space embeddings may introduce class-distance bias. (c) Our method maps degradations onto a unit hypersphere, enabling geometrically balanced distributions and robust branch selection via cosine similarity.