OpenNovelty.org

UniRestorer: Universal Image Restoration via Adaptively Estimating Image Degradation at Proper Granularity

ICLR 2026 Conference SubmissionAnonymous Authors
OpenReview Score: 6.0Download Report PDF
low-level visionimage restorationall-in-one image restoration
Abstract:

Recently, considerable progress has been made in all-in-one image restoration. Generally, existing methods can be degradation-agnostic or degradation-aware. However, the former are limited in leveraging degradation estimation-based priors, and the latter suffer from the inevitable error in degradation estimation. Consequently, the performance of existing methods has a large gap compared to specific single-task models. In this work, we make a step forward in this topic, and present our UniRestorer with improved restoration performance. Specifically, we perform hierarchical clustering on degradation space, and train a multi-granularity mixture-of-experts (MoE) restoration model. Then, UniRestorer adopts both degradation and granularity estimation to adaptively select an appropriate expert for image restoration. In contrast to existing degradation-agnostic and -aware methods, UniRestorer can leverage degradation estimation to benefit degradation-specific restoration, and use granularity estimation to make the model robust to degradation estimation error. Experimental results show that our UniRestorer outperforms state-of-the-art all-in-one methods by a large margin, and is promising in closing the performance gap to specific single-task models. The code and pre-trained models will be publicly available.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Overall Novelty Assessment

The paper introduces UniRestorer, a multi-granularity mixture-of-experts framework for all-in-one image restoration. It occupies the 'Hierarchical Granularity-Based Degradation Estimation' leaf within the taxonomy, which currently contains only this single paper. This indicates a relatively sparse research direction focused on multi-level degradation characterization. The broader parent category 'Degradation Representation and Guidance Mechanisms' contains six other papers across three sibling leaves, suggesting that while degradation-guided restoration is an active area, hierarchical granularity estimation represents a novel angle within this space.

The taxonomy reveals several neighboring approaches to degradation-aware restoration. The sibling leaf 'Prompt-Based Degradation Guidance' contains six papers using learned prompts or instructions for task conditioning, while 'Neural Degradation Encoding' includes three papers learning explicit degradation embeddings. The 'Task-Adaptive Architecture and Parameter Modulation' branch explores dynamic network configurations through hypernetworks and routing mechanisms. UniRestorer's hierarchical clustering of degradation space and granularity-based expert selection distinguishes it from these single-level prompt or encoding strategies, positioning it at the intersection of degradation estimation and adaptive architecture design.

Among the twenty-seven candidates examined through semantic search, the contribution-level analysis shows varied novelty profiles. The multi-granularity mixture-of-experts model examined seven candidates with zero refutations, suggesting this architectural choice is relatively distinctive. Joint degradation and granularity estimation examined ten candidates, also with no clear refutations. However, the overall UniRestorer framework examined ten candidates and found two potentially overlapping prior works, indicating that while specific technical components may be novel, the general approach of combining degradation estimation with adaptive expert routing has some precedent in the limited search scope.

Based on the top-27 semantic matches examined, UniRestorer appears to introduce a distinctive combination of hierarchical granularity estimation and mixture-of-experts routing. The sparse population of its taxonomy leaf and the absence of refutations for its core technical contributions suggest meaningful novelty in how it structures degradation space. However, the framework-level overlap with two candidates indicates that the broader concept of degradation-aware adaptive restoration is not unprecedented. The analysis reflects a focused literature search rather than exhaustive coverage of all-in-one restoration methods.

Taxonomy

Core-task Taxonomy Papers
50
3
Claimed Contributions
27
Contribution Candidate Papers Compared
2
Refutable Paper

Research Landscape Overview

Core task: all-in-one image restoration. The field addresses the challenge of restoring images degraded by diverse and often unknown distortions—such as noise, blur, rain, or haze—using a single unified model rather than task-specific networks. The taxonomy reveals several complementary research directions. Degradation Representation and Guidance Mechanisms focus on explicitly modeling or estimating degradation types and severities to steer restoration, with works like PromptIR[1] and Omni-Kernel[2] employing learnable prompts or kernel representations. Task-Adaptive Architecture and Parameter Modulation explores dynamic network configurations that adjust parameters or routing based on input characteristics, exemplified by AdaIR[31] and Adaptive Blind[4]. Generative Prior-Based Restoration leverages diffusion models and other generative priors, as seen in Diffusion Plug-and-Play[6] and Generative Diffusion Prior[9]. Meanwhile, Transformer and Attention-Based Architectures (e.g., Restormer[17], Uformer[12]) and Convolutional and State-Space Architectures (e.g., VmambaIR[5]) investigate foundational backbone designs. Additional branches address Task Balancing and Multi-Task Optimization, Blind and Adaptive Restoration for unknown degradations, and Specialized Applications in domains like medical imaging. A central tension across these branches is how to balance generality with specificity: some methods pursue fully blind restoration without explicit degradation cues, while others invest in fine-grained degradation estimation to guide the network more precisely. UniRestorer[0] sits within the Hierarchical Granularity-Based Degradation Estimation cluster, emphasizing multi-level degradation characterization to inform restoration decisions. This contrasts with approaches like All-In-One Unknown[3], which targets blind settings with minimal degradation priors, and PromptIR[1], which uses lightweight prompt learning for task conditioning. Compared to dynamic prompting strategies such as Dynamic Prompts[10] or instruction-based methods like InstructIR[14], UniRestorer[0] focuses on hierarchical granularity to capture both coarse degradation categories and fine-grained severity variations. This positioning reflects an ongoing exploration of how much and what kind of degradation information is necessary to achieve robust, generalizable all-in-one restoration across diverse real-world scenarios.

Claimed Contributions

Multi-granularity mixture-of-experts restoration model

The authors construct a hierarchical degradation representation set by clustering the degradation space at multiple granularity levels. They then train a corresponding multi-granularity MoE restoration model where fine-grained experts specialize in specific degradations while coarse-grained experts generalize across broader degradation spaces.

7 retrieved papers
Joint degradation and granularity estimation for adaptive expert routing

The method introduces granularity estimation alongside degradation estimation to indicate the degree of degradation estimation error. This joint estimation enables adaptive selection of the most appropriate expert, making the model robust to degradation estimation errors while leveraging degradation-specific restoration.

10 retrieved papers
UniRestorer framework for universal image restoration

The authors present UniRestorer, an all-in-one image restoration framework that estimates degradation at proper granularity levels and adaptively allocates experts. The framework significantly outperforms existing all-in-one methods and closes the performance gap with specific single-task models.

10 retrieved papers
Can Refute

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Within the taxonomy built over the current TopK core-task papers, the original paper is assigned to a leaf with no direct siblings and no cousin branches under the same grandparent topic. In this retrieved landscape, it appears structurally isolated, which is one partial signal of novelty, but still constrained by search coverage and taxonomy granularity.

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

Multi-granularity mixture-of-experts restoration model

The authors construct a hierarchical degradation representation set by clustering the degradation space at multiple granularity levels. They then train a corresponding multi-granularity MoE restoration model where fine-grained experts specialize in specific degradations while coarse-grained experts generalize across broader degradation spaces.

Contribution

Joint degradation and granularity estimation for adaptive expert routing

The method introduces granularity estimation alongside degradation estimation to indicate the degree of degradation estimation error. This joint estimation enables adaptive selection of the most appropriate expert, making the model robust to degradation estimation errors while leveraging degradation-specific restoration.

Contribution

UniRestorer framework for universal image restoration

The authors present UniRestorer, an all-in-one image restoration framework that estimates degradation at proper granularity levels and adaptively allocates experts. The framework significantly outperforms existing all-in-one methods and closes the performance gap with specific single-task models.

Table of Contents