All-in-One Image Restoration Survey: a living catalogue of degradation-aware models

A survey with a survey, and a repository with one file

The first thing to understand is that the repository and the paper are two things. The paper is titled A Survey on All-in-One Image Restoration: Taxonomy, Evaluation and Future Trends, and the README records it as an IEEE TPAMI 2025 item with an arXiv preprint at 2410.15067, submitted on 19 October 2024 as version one and last revised on 12 June 2025 as version two with what the news section calls major updates and reductions. The repository, by contrast, has a single top level entry, README.md, and nothing else. So the GitHub project is a living index that points at the paper rather than a software package, a dataset release, or a reimplementation of the models it lists. The authors are Junjun Jiang, Zengyuan Zuo, Gang Wu, Kui Jiang and Xianming Liu of the AIIA Lab at Harbin Institute of Technology, and the README states the intent to keep reflecting the latest work in the field.

The taxonomy is the intellectual contribution

The title's three parts describe the structure of the survey: taxonomy, evaluation and future trends. The taxonomy is the part that makes it more than a list of citations, because all-in-one image restoration is a crowded area where a single model is expected to handle several degradations at once, from rain and snow through haze and low light to noise, and the ways of doing that differ enough that a flat bibliography is not useful. The repository reflects that framing by grouping entries under a heading for the paper list and a separate heading for performance comparison, and by separating the method catalogue from the benchmark tables. If you are new to the area, that organisation is the point of the project. The trend section is where the newest directions surface, and the 2026 entries in the table already show a strong cluster of agentic work, including a paper on self-evolving agentic image restoration through deliberate planning and a VLM agent system for photographic post-processing.

The paper list runs year by year from 2020 to 2026

The paper list is organised by publication year, with anchors for 2020, 2021, 2022, 2023, 2024, 2025 and 2026, and each entry is a table row with a title, an avenue, a paper link and a code column. The venue column is what makes the list navigable rather than decorative, since it shows where the work actually appears: ICLR 2026 for RestoreVAR, AAAI 2026 for a degradation-aware LoRA approach, BMVC 2026 for continual learning across image degradations, NeurIPS 2026 for several entries including composite restoration, open-world adaptation and an adverse weather removal method, ECCV 2026 for the self-evolving agentic work, ACM MM 2026 for semantic-informed multimodal alignment, CVPR 2026 for lifelong learning in medical image restoration and for a unified VLM agent for post-processing, IJCV 2026 for video restoration and for generative diffusion priors, and TIP 2026 for adaptive frequency enhancement. Where a code repository exists the row links to it directly.

Benchmark tables grouped by dataset family

The second half of the repository is performance comparison, and its structure follows the datasets rather than the models. The navigation lists three challenging datasets, five challenging datasets, an all-weather section, a real-world weather-related section, the CDD-11 datasets, and a section for FoundIR's real-world datasets. That grouping matters for interpretation. A model that scores well on a synthetic rain dataset and badly on a real driving scene has learned the dataset rather than the task, and the real-world and all-weather tables are the ones that expose that gap. CDD-11 is listed as its own family, and the FoundIR datasets are attributed to that project, so the tables credit the data source rather than treating it as generic benchmark data. The tables are the most perishable part of the repository, since a result depends on the model version, the training data and the evaluation protocol as much as on the method.

Using the repository, and what it does not give you

There is nothing to install here, and the README gives no clone, build or run instructions, because there is nothing to run. The practical way to use the repository is to read it. The top navigation links to the paper list, to the benchmark tables, to a citations section and to a contributing section, and the paper list carries links out to the arXiv preprint, to the IEEE Xplore entry, to the authors and the lab, to the hits counter and to the awesome.re listing. The README also carries a badge linking to the survey on IEEE Xplore, which is the citable version if you need to reference the taxonomy rather than a blog post. What the repository does not give you is code. The code column in the paper table points at the individual authors' repositories, and several 2026 rows have an empty code cell, so if you are evaluating methods by whether you can run them, expect to drop a fair share of them.

What the entries say about where the field is going

Reading the 2026 rows together gives a picture of the current direction, and it is worth doing since the survey's own trend section is a claim you can check against the list. Three threads stand out. The first is efficiency, visible in a degradation-aware LoRA work presented at AAAI 2026 and in an adaptive frequency enhancement paper at TIP 2026, both of which treat the cost of serving one model for many degradations as a first class problem. The second is generalisation beyond known degradations, with a NeurIPS 2026 entry on discovering unseen degradations to adapt open-world image restoration and a CVPR 2026 entry on lifelong learning for medical image restoration, both of which ask what happens when the degradation at test time was never in the training set. The third is the agentic turn, with work on post-training GRPO, deliberate planning for self-evolving agents, and a VLM agent that fuses multiple experts rather than chaining tools sequentially.

Where the survey repository is the wrong tool

The limits are structural. First, it is a snapshot with a maintenance policy, not a live leaderboard, and the README's own news section gives the version history of the paper while the tables in the repository can be regenerated at any time, so the date of a number matters more here than it would in a paper. Second, a curated list inherits the biases of its curators, and a survey from a single lab with a single perspective on how the field should be decomposed will group things in a way that favours their own taxonomy. Third, for anyone who actually wants to restore an image, this repository is one or two clicks from the model repositories rather than being a tool itself, and the survey's value is orientation rather than capability. The alternative worth naming is reading the individual papers, which give the full experimental detail that a table row compresses away, or following the individual code repositories where the model weights and the training recipes live. The survey is most useful as the map you read first, before you decide which of the 2026 rows to open in full.

Editorial conclusion

Use this repository as a reading list and a benchmark index if you are starting work on all-in-one image restoration, or if you need a dated view of how the field splits into degradation-aware, prompt-based and agentic directions. Do not expect code from it, since the repository contains only a README, and do not treat the benchmark tables as settled rankings, because the results are snapshots of specific model versions on specific datasets. Verify first how recently the tables were regenerated, and check the individual paper entries for code links rather than assuming a method ships weights.

Frequently asked questions

Official sources

1. Harbinzzy/All-in-One-Image-Restoration-Survey on GitHub
2. Issues
3. README