Efficient Degradation-agnostic Image Restoration via Channel-Wise Functional Decomposition and Manifold Regularization
Bin Ren, Yawei Li, Xu Zheng, Yuqian Fu, Danda Pani Paudel, Hong Liu, Ming-Hsuan Yang, Luc Van Gool, Nicu Sebe
TL;DR
MIRAGE is a novel, efficient image restoration framework that uses channel-wise functional decomposition and manifold regularization to handle diverse corruptions with high performance and generalization.
Contribution
The paper introduces MIRAGE, a degradation-agnostic image restoration method combining channel-wise decomposition and manifold regularization for improved efficiency and generalization.
Findings
Achieves state-of-the-art performance across diverse degradation types.
Offers superior efficiency-performance trade-offs compared to existing methods.
Demonstrates strong generalization to unseen degradation scenarios.
Abstract
Degradation-agnostic image restoration aims to handle diverse corruptions with one unified model, but faces fundamental challenges in balancing efficiency and performance across different degradation types. Existing approaches either sacrifice efficiency for versatility or fail to capture the distinct representational requirements of various degradations. We present MIRAGE, an efficient framework that addresses these challenges through two key innovations. First, we propose a channel-wise functional decomposition that systematically repurposes channel redundancy in attention mechanisms by assigning CNN, attention, and MLP branches to handle local textures, global context, and channel statistics, respectively. This principled decomposition enables degradation-agnostic learning while achieving superior efficiency-performance trade-offs. Second, we introduce manifold regularization that…
Peer Reviews
Decision·ICLR 2026 Poster
1. The writing is clear and easy reading. 2. Experiments show the effectiveness of the proposed method.
1. The motivation of this work that using different strategy to handle agnostic degradations is much more invested in previous research, and the innovation is somewhat minimal; 2. There is no solid clarification that the shared representation between shallow and deep features are benefited for diverse degradations, which makes the contrastive objective not convincing, and the ablation improvement is minimal for probable performance jitter; 3. The main experiments in Tab. 3 and 4 may not show the
1. The SPD-space contrastive learning is a novel contribution that effectively preserves high-order statistics, outperforming Euclidean contrastive counterparts in both stability and generalization, as evidenced by ablation studies. 2. The channel-wise functional decomposition is conceptually clear and well-motivated by empirical redundancy analysis. It provides an elegant way to repurpose redundant channels into complementary submodules, rather than simply increasing model capacity. 3. The anal
1. The two main contributions claimed in the introduction are largely based on the combination and engineering integration of existing techniques, rather than presenting any genuinely novel innovation or introducing a fundamentally new perspective. 2. While the contributions are substantial, the paper is dense and may be challenging for readers unfamiliar with SPD manifolds or contrastive learning in feature covariance space. Some mathematical formulations could benefit from more intuitive expla
1. The paper presents a well-motivated and compact architecture that repurposes attention-channel redundancy through a channel-wise functional decomposition (Conv/Attn/MLP branches), aligning with the natural inductive biases of different degradations—local texture (Conv), global context (Attn), and channel statistics (MLP). 2. The proposed SPD manifold alignment introduces cross-layer contrastive learning between shallow and latent features using covariance-based second-order statistics. This
1. While the paper’s SPD-based manifold regularization is empirically effective, its mathematical formulation remains heuristic. The method projects SPD matrices back to Euclidean space for InfoNCE optimization, without leveraging true Riemannian metrics or geodesic distances. As a result, the claimed “manifold alignment” lacks formal theoretical justification. 2. The experiments comprehensively compare with transformer- and prompt-based all-in-one models (PromptIR, MoCE-IR, AdaIR), but omit rec
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsAdvanced Image Processing Techniques · Image and Signal Denoising Methods · Image Enhancement Techniques
MethodsSoftmax · Attention Is All You Need · Convolution · Contrastive Learning