M-3LAB/awesome-industrial-anomaly-detection

What this project is

M-3LAB/awesome-industrial-anomaly-detection is a curated, constantly updated list of research resources about a specific AI task: finding defects in manufactured products using images (called "industrial anomaly detection"). It's the kind of repository the AI research community calls an "awesome list" — a hand-picked collection of papers, benchmarks, and code, organized for easy browsing.

The project is maintained by M-3LAB, a research group. The README states they "discuss public datasets and related studies in detail" and welcome contributions (pull requests) from anyone who wants to add or reorganize papers.

What it contains

The repository is essentially a living index of the field, with several major parts:

• Survey papers — long-form overviews of the field. The README links to three: a general survey published in Machine Intelligence Research, a newer survey (2026) on recent advances, and a benchmark paper called IM-IAD (Industrial Image Anomaly Detection Benchmark in Manufacturing).
• SOTA methods with code — a table of ~25 notable research papers, each with its publication venue (e.g., CVPR, NeurIPS, ICCV), year, a link to the paper, a link to its open-source code, and a short topic label (e.g., "Teacher-Student", "Memory-bank", "Few Shot", "Multi-class unified").
• Recommended benchmarks — a smaller table of 5 benchmark tools/libraries, including Anomalib (a deep learning library for anomaly detection) and IM-IAD.
• Recent research by conference — papers grouped by where they were published (ECCV 2026, ICML 2026, CVPR 2026, ICLR 2026, AAAI 2026, NeurIPS 2025, KDD 2025, ICCV 2025, ICML 2025, CVPR 2025). Each entry has a title, a paper link, and often a code link.
• A paper tree and timeline — visual/structural classifications of representative methods and how the field evolved.
• Categorized paper lists — the bulk of the repository. Papers are grouped by research approach:
  • Unsupervised AD (methods that learn from only normal/defect-free images): feature-embedding-based (teacher-student, one-class classification, distribution-map, memory bank, vision-language), reconstruction-based (autoencoders, GANs, transformers, diffusion models), and supervised variants.
  • Other research directions: zero/few-shot AD, noisy AD, anomaly synthesis, RGBD AD, 3D AD, continual AD, multi-class AD, logical AD, MLLM-based AD, video IAD, and more.
• A dataset section — lists of public datasets used in this field.
• Citation info — a BibTeX entry so researchers can cite the repository.

Notable highlights mentioned in the README

• AD-Copilot — a "multimodal large language model" (MLLM) for industrial anomaly detection, which the README claims "surpasses humans on real industrial inspection tasks." It has a demo on Hugging Face.
• MMAD — a benchmark for evaluating how well current MLLMs perform as industrial quality inspectors (published at ICLR 2025).
• ASBench — a benchmark for comparing different methods of synthesizing artificial anomalies (for training).
• 3D anomaly detection — a separate curated list for point-cloud-based defect detection.

Who it's for

This is a research resource, not a software tool. It's aimed at:

• Researchers and graduate students working on industrial anomaly detection who need to find relevant papers, code, and benchmarks quickly.
• Anyone who wants to understand the landscape of this field — the categories show the main approaches and open problems.

How to use it

You browse it like a table of contents. If you're looking for, say, "few-shot anomaly detection methods," you'd go to the "3.1 Zero/Few-Shot AD" section. If you want a ready-to-run method, you'd look at the "SOTA methods with code" table. The README also encourages contributions — you can submit a pull request to add or re-categorize papers.

Key caveat

The README is a list of links and titles, not a tutorial. It doesn't explain how the methods work — it just points you to the papers that do. The actual learning happens by reading the linked papers and code.