Harbinzzy/All-in-One-Image-Restoration-Survey

[IEEE TPAMI 2025] A Survey on All-in-One Image Restoration: Taxonomy, Evaluation and Future Trends

What this project is

This GitHub repo is the companion to a published survey paper (in IEEE TPAMI) about all-in-one image restoration — AI models that take one image that has been damaged in various ways (rain, haze, noise, low light, etc.) and restore it to a clean version using a single unified model, instead of needing a separate model for each type of damage.

The repo itself is a continuously updated catalog: a list of 100+ research papers on this topic, organized by year (2020–2026), with links to each paper and its code. It also includes performance comparison tables — benchmark results for how well different all-in-one restoration models perform on standard test datasets (e.g., weather-related datasets, real-world datasets, a dataset called CDD-11, and FoundIR's real-world datasets).

The core promise

If you're a researcher or engineer trying to understand this fast-moving field, this repo lets you:

See every major paper on all-in-one image restoration from the last ~6 years, with one-row-per-paper entries that show the venue (e.g., CVPR, ICLR, NeurIPS), the direct paper link, and a code link when available.
Compare model performance across the standard benchmark datasets — you can look up tables that rank different restoration models on the same evaluation datasets.
Stay current — the repo explicitly says the survey is "actively maintained," and the "News" section notes it was first released Oct 2024 and last revised June 2025 (with major updates). The paper list even includes 2026 publications, which signals the maintainers keep adding new work as it appears.

Who made it and who it's for

Authors: Five researchers from the AIIA Lab at Harbin Institute of Technology (Junjun Jiang, Zengyuan Zuo, Gang Wu, Kui Jiang, Xianming Liu).
Primary audience: People working in computer vision / image processing — graduate students, researchers, and engineers who need an organized, citation-ready overview of the state of the art, plus benchmark numbers to evaluate their own models.

What's inside (concretely)

Section	What it contains
Paper List	Tables of publications, one per row: paper title, authors, venue (ICLR 2026, AAAI 2026, CVPR 2026, etc.), link to the paper (arXiv, IEEE, Springer, etc.), and link to the code repo when the authors released one. Grouped by year: 2020 → 2026.
Benchmark Tables	Performance comparisons of all-in-one restoration models, on: 3 challenging datasets, 5 challenging datasets, all-weather datasets, real-world weather-related datasets, the CDD-11 dataset, and FoundIR's datasets.
News	Announcement of the survey paper release and its revisions.
Navigation	Quick links to jump to each year or each benchmark section.

Interesting details worth noting

The repo badge says the paper is on arXiv (id 2410.15067) and also published in IEEE TPAMI (a top-tier journal in computer vision) — the badge reads "IEEE TPAMI 25".
Papers span venues like ICLR, AAAI, NeurIPS, ECCV, CVPR, ICML, ACM MM, TIP, TPAMI, IJCV, TNNLS, and workshops — so it's genuinely comprehensive across both conferences and journals.
Many entries include demo code links (GitHub URLs), but some don't — the README simply leaves the code column blank, so no code is claimed when none is provided.
The paper titles themselves reveal a clear trend: newer works (2025–2026) increasingly use diffusion models, agents / LLMs, and flow matching — the repo is effectively a timeline of how the field evolved.

How to use it

Start with the survey paper (linked via the arXiv badge) to get the full taxonomy, evaluation methodology, and future trends discussion.
Browse the paper list by year to find relevant work — each row links out to the original publication and code.
Compare benchmark tables to see which models score best on specific datasets.
Check the News section for updates — the maintainers promise continued revisions as the field evolves.

What it is NOT

It is not a code library or a framework you can run.
It is not a dataset download page.
It is a survey + curated bibliography + benchmark comparison — its value is in organizing and linking external resources, not in providing its own software.

All facts above are drawn directly from the README's content — the paper list entries, the navigation structure, author names, affiliations, publication badges, and benchmark section labels come verbatim from that document. The only inference I've added is labeling it a "curated bibliography" and "benchmark comparison" — which is what sections like "Paper List" and "Performance Comparisons" literally are.