Review Learning: Advancing All-in-One Ultra-High-Definition Image Restoration Training Method

TL;DR

This work tackles the challenge of universal UHD image restoration by introducing Review Learning, a prompt-free continual-learning paradigm that sequentially trains on degraded datasets while periodically rehearsing challenging samples to mitigate forgetting. The authors pair this training strategy with SimpleIR, a lightweight all-in-one restoration network featuring Hybrid Attention Blocks, a Local Detail Attention Module, and Feature Iteration Blocks to enable full-resolution 4K inference on consumer GPUs. Across LLIE, deraining, desnowing, and deblurring tasks, SimpleIR achieves competitive or state-of-the-art results with fewer parameters, and maintains strong performance in multi-task settings, validated by ablations on review quantity and learning trajectories. The approach eliminates reliance on prompts or bespoke architectures, offering a practical, scalable solution for UHD restoration in real-world pipelines and downstream tasks such as object detection.

Abstract

All-in-one image restoration tasks are becoming increasingly important, especially for ultra-high-definition (UHD) images. Existing all-in-one UHD image restoration methods usually boost the model's performance by introducing prompt or customized dynamized networks for different degradation types. For the inference stage, it might be friendly, but in the training stage, since the model encounters multiple degraded images of different quality in an epoch, these cluttered learning objectives might be information pollution for the model. To address this problem, we propose a new training paradigm for general image restoration models, which we name \textbf{Review Learning}, which enables image restoration models to be capable enough to handle multiple types of degradation without prior knowledge and prompts. This approach begins with sequential training of an image restoration model on several degraded datasets, combined with a review mechanism that enhances the image restoration model's memory for several previous classes of degraded datasets. In addition, we design a lightweight all-purpose image restoration network that can efficiently reason about degraded images with 4K ($3840 \times 2160$) resolution on a single consumer-grade GPU.

Figures (11)

• Figure 1: This paper leverages the Review Learning training method for an all-in-one image restoration task. In contrast to existing all-in-one modeling methods for image restoration, we only leverage customized training schemes to optimize the image restoration network without a priori and prompting information. Our method achieves insightful performance on seven benchmarks (including degraded UHD images and low-resolution images) compared to other algorithms.
• Figure 2: We propose a review learning method that trains an end-to-end network sequentially on different types of datasets. It is worth noting that to avoid catastrophic forgetting of the model, our network is trained on a new dataset by reviewing previously trained difficult samples of other degenerate types. These difficult samples are stored by observing fluctuations in the loss during training.
• Figure 3: Histograms of entropy differences for 4 image degradation datasets.
• Figure 4: The training loss curve illustrates the model's performance across various stages of multi-degradation task learning.
• Figure 5: Qualitative comparisons with NAFNet, Restormer, PromptIR, and AirNet on LLIE, deblurring, deraining, and desnowing tasks. SimpleIR can generate cleaner results with finer details. See the Appendix for more results. (Zoom in for a better view)