Bridging Degradation Discrimination and Generation for Universal Image Restoration
JiaKui Hu, Zhengjian Yao, Lujia Jin, Yanye Lu
TL;DR
This paper introduces BDG, a novel universal image restoration method that combines degradation discrimination and generation, improving performance across multiple degradation types without changing architecture.
Contribution
The paper proposes MAS-GLCM for fine-grained degradation discrimination and a three-stage diffusion training process to enhance restoration capabilities.
Findings
Significant performance improvements in all-in-one restoration tasks
Enhanced fidelity in real-world super-resolution
Effective discrimination of degradation types and levels
Abstract
Universal image restoration is a critical task in low-level vision, requiring the model to remove various degradations from low-quality images to produce clean images with rich detail. The challenges lie in sampling the distribution of high-quality images and adjusting the outputs on the basis of the degradation. This paper presents a novel approach, Bridging Degradation discrimination and Generation (BDG), which aims to address these challenges concurrently. First, we propose the Multi-Angle and multi-Scale Gray Level Co-occurrence Matrix (MAS-GLCM) and demonstrate its effectiveness in performing fine-grained discrimination of degradation types and levels. Subsequently, we divide the diffusion training process into three distinct stages: generation, bridging, and restoration. The objective is to preserve the diffusion model's capability of restoring rich textures while simultaneously…
Peer Reviews
Decision·ICLR 2026 Poster
1. BDG introduces the Multi-Angle and multi-Scale Gray Level Co-occurrence Matrix (MAS-GLCM), which significantly improves the model’s ability to distinguish between various degradation types and levels. This enables more precise restoration tailored to the input image’s condition. 2. By aligning MAS-GLCM features with diffusion model features during a dedicated bridging stage, BDG effectively combines degradation awareness with generative priors. This fusion allows the model to retain rich text
1. While MAS-GLCM is presented as a novel degradation discriminator, it is fundamentally an extension of the classical Gray Level Co-occurrence Matrix (GLCM), which has been widely used in texture analysis for decades. To strengthen the novelty claim, the authors should compare MAS-GLCM with frequency-aware representations and learned degradation embeddings (e.g., from PromptIR or DCPT) or (e.g., Ji et al., 2021). Ablation studies showing MAS-GLCM’s superiority over these alternatives would hel
The paper achieves impressive empirical results, particularly in reconciling high fidelity metrics (PSNR) with the strong perceptual quality of diffusion models in real-world super-resolution .
My primary concern lies with the methodological innovation. The core degradation descriptor, MAS-GLCM , is a classical, handcrafted feature descriptor. While its effectiveness in separating degradations is well-demonstrated (Fig. 1), this approach feels like a step back from end-to-end deep feature learning. The system relies on this external, non-learned feature extractor, which is then 'aligned' with the diffusion U-Net's features. This raises questions about whether a more integrated, end-to-
+ The unified formulation connecting discrimination and generation is interesting. + Achieving quantitative and qualitative improvements over recent SOTA (DiffUIR, DCPT).
1. Though losses and stages are ablated, there is no comparison of MAS-GLCM vs. simpler features (e.g., gradients, frequency) in restoration quality. 2. The paper does not report details about computational cost (params / memory overhead / time) of MAS-GLCM. 3. The number of scales × angles used ($l$, $\theta$) and preprocessing details are not discussed. 4. Questions about MAS-GLCM : (1) Although the authors claim MAS-GLCM is “minimally affected by image content,” GLCM is a gray-le
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsAdvanced Image Processing Techniques · Image and Video Quality Assessment · Image Enhancement Techniques