Discovering Unseen Degradations
to Adapt Semi-Supervised Image Restoration

While all-in-one image restoration models excel in controlled, closed-set environments, they face critical limitations when deployed in open-world settings where paired supervision is unavailable and degradations are unknown and often intricately mixed. To address this challenge, we propose a continual restoration framework that integrates degradation discovery with adaptive restoration. Building on a model pre-trained on known degradation categories, our method identifies unseen degradations by measuring deviations in feature space via Mahalanobis distance and performs continual category discovery to expand the degradation space. Based on this, we introduce a discovery- and instance-conditioned descriptor that combines degradation-aware representations derived from cluster prototypes with instance-specific features, enabling adaptive conditioning for mixed and ambiguous degradations. To further handle unpaired data, we adopt a mean-teacher-based semi-supervised framework with a reliable bank for pseudo-target refinement. Crucially, we propose a discovery-adaptive score that evaluates candidate targets based on their relative proximity to degradation clusters and an anchored clear cluster, enabling degradation-aware pseudo-target selection beyond conventional no-reference image quality assessment metrics. Experiments on both controlled open-world protocols and real-world adverse weather datasets demonstrate that our approach consistently outperforms existing methods across multiple image quality metrics while improving robustness to degradation composition shifts.

Security cameras and autonomous cars often capture hard-to-see images in rain, snow, fog or darkness. AI tools can fix these, but are usually trained on a fixed list of problems, using pairs of damaged and clear photos of the same scene. Outdoors, new and mixed conditions keep appearing, and such pairs rarely exist: no one can photograph a street in storm and sunshine at once.

We present DUDA, which keeps learning on the job without clear photos or human labels. Like a birdwatcher who starts a new page for a bird missing from the guidebook and adds later sightings to it, DUDA sorts images by type of damage, adding new types as they appear. These discoveries guide each repair, and DUDA learns only from its most trustworthy results.

This could keep outdoor cameras reliable as conditions change. In tests, its restored images also helped detectors spot cars and people better.