UniRestorer: Universal Image Restoration via Adaptively Estimating Image Degradation at Proper Granularity
Jingbo Lin, Zhilu Zhang, Wenbo Li, Renjing Pei, Hang Xu, Hongzhi Zhang, Wangmeng Zuo
TL;DR
UniRestorer introduces a hierarchical, multi-expert image restoration model that adaptively estimates degradation and granularity, significantly improving performance over existing methods and narrowing the gap with specialized single-task models.
Contribution
It proposes a novel multi-granularity mixture-of-experts model with adaptive degradation and granularity estimation for improved all-in-one image restoration.
Findings
Outperforms state-of-the-art all-in-one methods by a large margin.
Leverages degradation and granularity estimation for robust, degradation-specific restoration.
Effectively narrows the performance gap to single-task models.
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-specific restoration, 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…
Peer Reviews
Decision·ICLR 2026 Poster
1. The paper proposes UniRestorer, the first framework that simultaneously exploits degradation and granularity estimation to overcome the inherent limitations of both degradation-agnostic and degradation-aware restoration methods. 2. Extensive quantitative and qualitative experiments convincingly demonstrate the superiority of UniRestorer over existing all-in-one baselines and its competitive performance against task-specific models. 3. The idea of granularity-aware expert selection is clearly
1. The paper lacks a quantitative analysis of the proposed hierarchical degradation-clustering step. 2. No comparison or discussion is provided against alternative clustering strategies (e.g., the spectral clustering adopted in SEAL). 3. It is unclear whether the Restormer baseline in Table 1 was re-trained under exactly the same degradation protocol and parameter budget; an ablation that removes both degradation and granularity estimation while keeping the backbone capacity fixed would better i
Granularity estimation elegantly handles degradation estimation noise — robust and practical. Hierarchical clustering + MoE scales well across 15+ degradation types. Large, consistent gains (e.g., +2.1 dB PSNR on mixed test sets). Comprehensive ablations (granularity levels, routing loss, expert count). Clean figures: t-SNE of degradation space, expert activation heatmaps.
Clustering is offline and static — no online adaptation to new/unseen degradations. Granularity estimator adds overhead — no inference latency reported (vs. PromptIR, AirNet). No theoretical justification for hierarchical clustering choice (e.g., why 3 levels?). Evaluation limited to synthetic degradations — no real-world camera pipeline (e.g., RAW → ISP). MoE training unstable? No mention of load balancing loss or expert collapse.
1. This paper introduces a multi-granularity degradation representation that unifies coarse- and fine-grained experts. The idea of granularity estimation to quantify degradation uncertainty and guide routing is novel and intuitive. 2. Comprehensive experiments: covers 7 single-degradation and 11 mixed-degradation settings, plus real-world and unseen tasks. 3. Paper is well-organized and technically detailed with intuitive figures (especially Fig. 3 illustrating routing). 4. The hierarchical MoE
1. While granularity estimation is conceptually convincing, there’s no formal uncertainty-theoretic or probabilistic analysis of its behavior. 2. Each granularity level adds parameters and routing complexity; actual FLOPs and latency comparisons are limited. The LoRA variant helps, but trade-offs between full and LoRA experts could be better quantified. 3. The K-means–based clustering assumes a consistent degradation embedding space; sensitivity to clustering hyperparameters (number of clusters,
Code & Models
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