Adaptive Blind All-in-One Image Restoration

TL;DR

This paper addresses blind all-in-one image restoration by proposing ABAIR, a three-phase framework that generalizes to unseen and composite degradations while allowing efficient updates. It combines large-scale synthetic pretraining with Degradation CutMix and a per-pixel degradation segmentation head, task-specific LoRA adapters, and a lightweight degradation estimator to blend adapters on a per-image basis. The approach outperforms state-of-the-art methods on five-task and three-task IR benchmarks and demonstrates robust performance on unseen datasets and mixed degradations, with minimal additional training when adding new degradations. The results indicate strong practical potential for versatile IR in real-world conditions, where degradations are diverse and evolving.

Abstract

Blind all-in-one image restoration models aim to recover a high-quality image from an input degraded with unknown distortions. However, these models require all the possible degradation types to be defined during the training stage while showing limited generalization to unseen degradations, which limits their practical application in complex cases. In this paper, we introduce ABAIR, a simple yet effective adaptive blind all-in-one restoration model that not only handles multiple degradations and generalizes well to unseen distortions but also efficiently integrates new degradations by training only a small subset of parameters. We first train our baseline model on a large dataset of natural images with multiple synthetic degradations. To enhance its ability to recognize distortions, we incorporate a segmentation head that estimates per-pixel degradation types. Second, we adapt our initial model to varying image restoration tasks using independent low-rank adapters. Third, we learn to adaptively combine adapters to versatile images via a flexible and lightweight degradation estimator. This specialize-then-merge approach is both powerful in addressing specific distortions and flexible in adapting to complex tasks. Moreover, our model not only surpasses state-of-the-art performance on five- and three-task IR setups but also demonstrates superior generalization to unseen degradations and composite distortions.

Figures (18)

• Figure 1: Our model significantly outperforms state-of-the-art all-in-one image restoration (IR) methods, DiffUIR zheng2024selective, X-Restormer chen2023x-restormer, and AdaIR cui2025adair, across five known IR tasks, three unseen tasks, and three mixed degradation scenarios. The plot is normalized along each axis, with the lowest value positioned on the second circle and the highest value on the outermost circle.
• Figure 2: General schema of our proposed method. Our method is divided into three phases. In Phase I we pretrain our baseline with synthetic degradations on high-fidelity images. Each image contains different degradations in different regions --- Degradation CutMix yun2019cutmix, and a segmentation head learns to predict them, while a restoration loss aims at restoring the image. In this way, the model can distinguish and generalize well to multiple degradations in a controlled manner. In Phase II, we learn degradation-specific adaptors using standard image restoration datasets. In Phase III, we learn a lightweight degradation estimator to adaptively blend the adapters based on the degradation profile of the input image. This three-phase methodology makes our method flexible to deal with images containing multiple degradations and easy to update for new ones as it only requires training an adapter for the new distortion and retraining the lightweight estimator.
• Figure 3: Rain
• Figure 4: Haze
• Figure 5: Noise