UniRestorer++: Scalable Universal Image Restoration with Multi-Granularity Degradation Modeling
Jingbo Lin, Zhilu Zhang, Fan Li, Renjing Pei, Wenbo Li, Hang Xu, Hongzhi Zhang, Wangmeng Zuo
Published November 4, 2026 · on the feed November 5, 2026 · ▲ 2 · arXiv 2611.06042
The abstract
Universal image restoration aims to handle diverse degradations within a single model, but a reliable scaling strategy remains underexplored. Increasing backbone depth and width can lead to performance saturation, while degradation-aware specialization depends on accurate degradation estimation and is vulnerable to errors on out-of-distribution and real-world inputs. We propose UniRestorer++, a scalable framework based on multi-granularity degradation modeling. It progressively refines degradation groups and introduces group-specific experts, coupling capacity growth with specialization. Degradation-estimation-based routing selects a root-to-leaf path, while granularity-estimation-based routing chooses a level along that path to balance specialization and robustness. Spatially dense degradation representation maps further enable region-aware routing for spatially non-uniform corruptions. We develop CDD-50, a benchmark covering 50 degradation categories from single to penta-degradation. Experiments show approximately log-linear PSNR trends over the evaluated model-size range and stronger scaling gains on out-of-distribution and real-world data than vanilla MoE scaling. UniRestorer++ achieves state-of-the-art performance across benchmarks and demonstrates zero-shot generalization and few-shot transfer on real-world images. Code: https://github.com/mrluin/UniRestorer_plusplus
The Brief
1:51A short podcast summary of the paper. Written and voiced by AI.
Transcript
UniRestorer++. Scalable Universal Image Restoration with Multi-Granularity Degradation Modeling. By Jingbo Lin and seven others. Published on the fourth of November.
Universal image restoration asks one model to undo many kinds of degradation, and how to scale such a model reliably is still an open matter. The two obvious routes each break. Making the backbone deeper and wider can run into performance saturation. Specializing by degradation type depends on estimating the degradation accurately, and that estimate is vulnerable to errors on out-of-distribution and real-world inputs. One route stops paying, and the other is fragile exactly where it is needed most.
UniRestorer++ ties the two together through multi-granularity degradation modeling. It progressively refines degradation groups and introduces experts specific to each group, so that adding capacity and adding specialization happen as one step rather than separately.
Two kinds of routing decide which expert handles an input. Routing based on degradation estimation selects a path from the root of the group hierarchy to a leaf. Routing based on granularity estimation then chooses a level along that path, balancing specialization against robustness. Spatially dense maps of the degradation representation extend this to region-aware routing, for corruptions that are not uniform across the image. To test all of this, the authors also build CDD-50, a benchmark covering fifty degradation categories, from single degradations up to penta-degradation.
Over the range of model sizes evaluated, PSNR follows an approximately log-linear trend. Scaling gains on out-of-distribution and real-world data are stronger than those of vanilla MoE scaling. UniRestorer++ is reported as state of the art across benchmarks, with zero-shot generalization and few-shot transfer on real-world images.
Its answer to the scaling question is to grow capacity along the structure of the degradations themselves, and to let the model decide how specific it can afford to be on each input.
More with an account
Listen to every day’s papers
Daily playlists, deep dives, quests for the papers, and quizzes on their key terms.
Join free