Anomalib
github.com
- Free plan
- Open source
- Windows
- Linux
- Web
- API
- Self-hosted
What Anomalib is
Anomalib is a free, Apache-2.0-licensed deep learning library for developing, benchmarking, and deploying anomaly detection algorithms focused on images and videos. Its modular Python API and command-line interface support training, inference, and benchmarking, and the project includes ready-to-use algorithms and benchmark datasets. Model implementations are based on Lightning, with inference options including Torch, Lightning, Gradio, and OpenVINO. Most models can be exported to OpenVINO Intermediate Representation for accelerated inference on Intel hardware. Documented experiment-tracking integrations include Weights & Biases, Comet.ml, and TensorBoard through PyTorch Lightning loggers. Installation options cover CPU, NVIDIA CUDA on Linux or Windows, AMD ROCm on Linux, and Intel XPU on Linux. Anomalib Studio is a low/no-code web application that accepts USB or IP cameras or image folders and can output to industrial pipelines through ROS messages or MQTT. Studio is a pre-release offered as a Docker container or standalone application, and its features may change or be incomplete or unstable.
Who it suits
Anomalib suits developers and teams building image or video anomaly detection workflows. Studio may suit users who want a low/no-code interface, provided they can work with pre-release software.
Good to know, and what to watch
- Free library under the Apache-2.0 license.
- Includes algorithms and benchmark datasets.
- Supports several inference options, including OpenVINO.
- Documents integrations with three experiment-tracking tools.
- Studio accepts camera feeds or image folders.
- Studio is a pre-release and may be incomplete or unstable.
- Intel GPU training currently supports only a single GPU.
- AMD ROCm installation is listed for Linux only.
Plans
Every plan and priceOpen-source Anomalib library
Open-source library
Open source
Getting started
- Install the open-source library from PyPI or source
- Select CPU, CUDA, ROCm or Intel XPU according to your hardware
- Use the Python API or CLI for training, inference or benchmarking
- Choose Torch, Lightning, Gradio or OpenVINO for inference
- Connect documented logging integrations if experiment tracking is needed
- Run Anomalib Studio as a Docker container or standalone application
Limits to check
Anomalib Studio is a pre-release, and its functionality may be incomplete, unstable or changed during development. Intel GPU training currently supports only a single GPU.
The Club's review
Overview
Anomalib is a Python library for developing and benchmarking deep-learning systems that detect or localize anomalies in images and videos. It is best suited to developers and technical teams building their own inspection or monitoring workflows. Its broad model and inference toolkit is compelling for that work, though it is not a ready-made detection service.
Key features
A modular Python API and command-line interface cover training, inference and benchmarking. Ready-to-use algorithms and benchmark datasets give teams a basis for comparing approaches and adapting them, rather than requiring every experiment to start from scratch. The models are built on Lightning, and integrations with Weights & Biases, Comet.ml and TensorBoard support experiment logging through PyTorch Lightning loggers.
Inference can run through Torch, Lightning, Gradio or OpenVINO. Most models can also be exported to OpenVINO Intermediate Representation for accelerated inference on Intel hardware. That gives teams targeting Intel devices a useful deployment route, but it should not be read as a promise of acceleration across hardware platforms.
Anomalib Studio provides a low/no-code web interface for working with USB or IP cameras and image folders. It can send outputs to industrial pipelines through ROS messages or MQTT, which may help teams explore a camera-based workflow without building every interface themselves. Studio is a pre-release under active development: features may change, and functionality may be incomplete or unstable. It is offered as a Docker container or standalone application, but production users should weigh that maturity caveat carefully.
Pricing
The Open-source library plan costs nothing and is licensed under Apache-2.0. It can be installed from PyPI or source. This makes it accessible for individual developers and teams that can supply their own development and deployment infrastructure; there is no paid tier to unlock in the plan described.
Platforms
Anomalib supports Linux and Windows, as well as web and self-hosted use. Hardware installation options include CPU, CUDA on Linux or Windows with NVIDIA GPUs, ROCm on Linux with AMD GPUs, and Intel XPU on Linux. Intel GPU training currently supports only one GPU; the project notes testing on Arc 750 and Arc 770. That single-GPU limit matters to teams planning larger Intel GPU training runs.
Who it's for
Anomalib fits developers, computer-vision practitioners and engineering teams that want to compare anomaly-detection algorithms, train models, and choose an inference path for image or video data. It is particularly relevant where Intel hardware is part of the deployment target, or where a team wants to connect a visual workflow to ROS or MQTT. Organizations seeking a turnkey service, or a stable low-code production interface, should look elsewhere or treat Studio as an experiment rather than a dependable core system.
Pros and cons
- Pros: A single library brings training, inference and benchmarking together with algorithms and datasets, giving developers a practical foundation for evaluation.
- Pros: Torch, Lightning, Gradio and OpenVINO inference options leave room to fit deployment to a team's existing stack, while OpenVINO export serves Intel-focused deployments.
- Pros: Apache-2.0 licensing and a free library make it possible to adopt without a license fee.
- Cons: Anomalib is a development toolkit, so teams must build and operate the workflow around its models rather than rely on a finished detection service.
- Cons: Studio's pre-release status means its feature set and stability may change, making it a riskier choice for production-critical use.
- Cons: Intel GPU training is currently limited to a single GPU, constraining that training setup.
Alternatives
For related options, browse Anomaly Detection Software.
- Metaplane is a freemium web and API option with a free tier capped at 10 monitored tables, 4 users and 3 custom SQL monitors; consider it when those limits suit your needs.
- OpenSearch is another free, Apache 2.0-licensed option that runs across API, desktop, web and self-hosted platforms; choose it when those platform options are a better fit.
- ManageEngine NetFlow Analyzer offers a free edition limited to 2 interfaces and paid Standard and Professional editions; consider it for a freemium alternative.
- ObservabilityOS has a free developer plan capped at 1 service, 500MB of logs per month and 7-day retention; it may suit a small observability setup within those caps.
- PyOD is a free, open-source Python library for API, Linux, macOS, self-hosted and Windows use; choose it when its library offering better fits your environment.
- Soda offers a free tier with processing units, pipeline testing, metrics observability and alerting and ticketing integrations; consider it for that data-quality workflow.
- Axomaly offers a free alpha with one metric, guided onboarding and direct support; it may suit readers seeking that guided starting point.
- PySAD is a free, open-source Python framework for Linux, macOS and Windows; consider it as another framework option.
Verdict
Choose Anomalib if you need a free, Apache-2.0 toolkit for developing and benchmarking image or video anomaly detection, and want flexibility across model training and inference. Its range of deployment options and Intel support are the strongest reasons to pick it. Look elsewhere if you need a complete detection service or a production-ready Studio experience.
The verdict
Anomalib provides a broad toolkit for developing and deploying image and video anomaly detection, with multiple inference and hardware options. Studio adds a visual workflow, but it remains under active development and may change.
Questions people ask
What license does Anomalib use?
The repository identifies the library as Apache-2.0 licensed.
Which data types can it analyze?
Its anomaly-detection focus covers images and videos.
Where can it run?
Platforms include Linux, Windows, web and self-hosted deployments. Hardware installation options include CPU, NVIDIA CUDA, AMD ROCm and Intel XPU as listed.
What is Anomalib Studio?
Studio is a low/no-code web application that accepts USB or IP cameras or image folders and can output through ROS messages or MQTT.
How can the library be installed?
The open-source library can be installed from PyPI or from source.
Which experiment trackers are documented?
Documented integrations include Weights & Biases, Comet.ml and TensorBoard through PyTorch Lightning loggers.
What the maker publishes
- Algorithm collection
- The documentation lists ready-made image models including PatchCore, PaDiM, DRAEM, FastFlow, GANomaly, Reverse Distillation, STFPM and WinCLIP. source
- Algorithms and datasets
- The project describes its collection as ready-to-use deep learning anomaly detection algorithms and benchmark datasets. source
- API and CLI
- It provides a modular Python API and command-line interface for training, inference, benchmarking and hyperparameter optimization. source
- Breaking changes
- The 2.7.0 changelog removes the legacy ONNX exporter, several deprecated APIs and older checkpoint migration support. source
- Core dependency requirement
- The v2.7.0 README badge states minimum PyTorch 2.10+, Lightning 2.2+ and OpenVINO 2024.0+. source
- Datasets
- Anomalib includes datamodules for image, video and depth datasets and documents public benchmark datasets. source
- Deployment
- Inference options include Torch, Lightning, Gradio, and OpenVINO. source
- Detection scope
- The project targets anomaly detection, anomaly localization and anomaly segmentation for images and videos. source
- Documentation depth
- The documentation includes reference APIs, CLI, data and model guides, inference, metrics, pipelines, tutorials and developer guides. source
- Edge inference
- Most models can be exported to OpenVINO Intermediate Representation for accelerated inference on Intel hardware. source
- Experiment tracking
- The documented logging integrations include Weights & Biases, Comet.ml, and TensorBoard through PyTorch Lightning loggers. source
- Experiment tracking
- Optional logging integrations include Weights & Biases, Comet.ml, TensorBoard and MLflow. source
- Hardware extras
- Installation extras cover CPU, NVIDIA CUDA 12.6 or 13.0, AMD ROCm and Intel XPU backends. source
- Hardware support
- Installation options include CPU, CUDA on Linux or Windows with NVIDIA GPUs, ROCm on Linux with AMD GPUs, and Intel XPU on Linux. source
- Independent distribution evidence
- SourceForge hosts an unaffiliated mirror and lists zero downloads for the displayed release folders. source
- Independent package status
- PyPI classifies Anomalib 2.7.0 as production/stable software for developers and scientific or engineering use. source
- Independent research description
- The peer-reviewed project paper describes Anomalib as an open-source, plug-and-play library with experiment trackers, visualizers, hyperparameter optimizers and OpenVINO optimization. source
- Industrial outputs
- Studio can send outputs into industrial pipelines through ROS messages and MQTT. source
- Inference options
- Inference support covers Torch, Lightning, Gradio and OpenVINO workflows. source
- Installation
- The recommended standard installation is through PyPI using pip or uv, with source installation available for development. source
- Intel GPU limit
- The README says Intel GPU training currently supports only a single GPU and notes testing on Arc 750 and Arc 770. source
- Intel GPU limit
- Training on Intel GPUs currently supports only a single GPU and was tested on Arc 750 and Arc 770. source
- Licence
- The repository and package metadata identify Apache License 2.0 as the project licence. source
- License
- The repository identifies its license as Apache-2.0. source
Checked against the maker's pages on 15 Oct 2026.