VAND 4.0

Challenge at CVPR'26

Welcome to the Visual Anomaly and Novelty Detection 2026 Challenge, VAND 4th Edition workshop at CVPR 2026!

Our workshop challenge aims to showcase current progress in anomaly detection across different practical settings while addressing critical issues in the field. Building on the encouraging results from previous years — including the VAND 3.0 challenge — this edition sets its sights even higher, pushing the boundaries of robust and generalizable anomaly detection models for real-world use cases, for the first time including both industrial and retail logistics focused competitions.

Challenge Objective and Tracks

Prizes

Important Dates

VAND 4.0 Official Challenge Rules

VAND 4.0 Industrial Track - Technical Guidelines

VAND 4.0 Retail Track - Technical Guidelines

Challenge Organizers

Challenge Objective and Tracks

The challenge addresses critical industrial needs for reliable anomaly detection under varying conditions and with limited data. We aim to bridge academic research with industrial requirements to develop solutions directly applicable to manufacturing, retail logistics, and beyond. The challenge hosts two individual tracks:

Industrial Track: MVTec AD 2 (go to leaderboard)
Retail Track: Kaputt 2 (go to leaderboard)

Participants can choose a single track or enter both in two separate submissions. These challenge tracks aim to advance existing anomaly detection literature and increase its adaptation in real-world settings. We invite the global community of innovators, researchers, and technology enthusiasts: Engage with these challenges and contribute towards advancing anomaly detection technologies in real-world scenarios.

From April 1st to May 14th (AOE), 2026, the global community will showcase its ideas on how to solve these challenges in visual anomaly detection.

You will find a detailed description of the individual tracks below, explaining the datasets used, the requirements for model design, the evaluation protocol, and points of contact, respectively. When in doubt about the participation and submission guidelines, please do not hesitate to reach out to us.

Prizes

Participate for a chance to win prizes!

Apart from model performance, computational efficiency and innovative method design will be rewarded.
For complete details on prize categories, eligibility criteria, and award amounts, please refer to the official VAND 4.0 Challenge rules.

Prize Categories per Track

Best Performance: Submission with overall best performance. ($3000 USD, prize provider: Amazon)
Best Performance runner-up: Submission with the overall second best performance. ($750 USD, prize provider: Voxel51)
Best Efficiency: Submission with best efficiency-performance trade-off. (appr. $6000 USD, prize provider: Intel)
Best Performance Zero-Shot/Off-the-shelf VLM: Submission with overall best performance under the Zero-Shot (Industrial Track)/ Off-the-shelf VLM (Retail Track) Setting. ($3000 USD, prize provider: Amazon).
Best Paper: Jury prize for an outstanding idea or methodological design. (appr. $400 USD, prize provider: MVTec)

General Aspects

Important Dates

Release of Detailed Track Descriptions: March 25th
Challenge time: April 1st – May 14th (Anywhere on Earth)
Submission Deadline: May 14th 23:59pm AOE
Results Announcement: May 21st
VAND4.0 workshop @CVPR26: June 3rd/4th
Winners will have the chance to present their solution as a short video (5 min).
Submission deadline for the video: May 28th (video presentation of the solution required to be eligible for prizes)

Model development

Participants are encouraged to explore and leverage any state-of-the-art anomaly detection, machine learning and vision(-language) models without limitations. Creativity and originality in model architecture and training methodology are strongly encouraged as long as they fulfil the track-specific requirements.

Datasets

For the Industrial Track, the MVTec AD 2 dataset will be used. Its design allows for evaluating models under real-world distribution shifts induced by changes in lighting conditions.

For the Retail Track, the Kaputt2 dataset will be used. The dataset is an extension of the Kaputt dataset released in 2025, illustrating scientific challenges for defect detection arising in retail logistics scenarios such as heavy pose and appearance variation. Kaputt2 will be released and made available at challenge start (April 1st).

Resources

Participation Guidelines

This challenge is open to individuals, teams, and academic and corporate entities worldwide. Submission requirements, evaluation metrics, and additional eligibility criteria — in particular with respect to prize eligibility — are detailed in the official rules which can be downloaded below.

VAND 4.0 Official Challenge Rules

Official Challenge Rules are available for download via the following link:
https://drive.google.com/file/d/1PWFG6QLaDV3ijw6HWB7rTLKiENrJwTvU/view?usp=sharing

Code Template

Please follow the structure of this template in your code repository to ensure resproducibility:

https://github.com/cvpr-vand/vand-2026

Report Template

Please follow the structure of this template in your technical report: https://docs.google.com/document/d/1mM4KGwZhxpv_n4Cv_63KshXNOe294AvTQaayn-YX-5w/edit?usp=drive_link

Detailed Track Information

VAND 4.0 Industrial Track - Technical Guidelines

Technical Guidelines for the Retail Track are availbale for download via the following link:
https://drive.google.com/file/d/1pVfl5NXB0WldFLeNQUvEdnv7dsGU3Wd5/view?usp=sharing

Please expand this section to see answers to Frequently Asked Questions about the Industrial Track.

VAND 4.0 Industrial Track - FAQs

We are a team of multiple participants. Do we collectively share the same 3 submissions per week, and what happens if different team members attempt submission?
You collectively share 3 submissions per week as a team. In case you create multiple accounts, please also be sure not to violate this limit.

Can I use pretrained backbones/weights such as DINOv3?
Yes, as long as these are easily publicly accessible.
Please consider the respective license terms when including pretrained backbones/weights into your model checkpoints.

What does “class-agnostic” mean in the context of this challenge?

A class-agnostic approach means that your method must follow a unified design across all object categories. The pipeline should not rely on manually tuned, category-specific decisions. Instead, any adaptation to a category must happen automatically based on normal data or predefined rules, not manual intervention.

Does the class-agnostic requirement apply to the entire pipeline?

Yes. The requirement applies to the entire pipeline, including:

Preprocessing
Prompt generation (if applicable)
Model architecture and inference
Thresholding
Postprocessing

All components should avoid manual per-category customization.

Are category-specific prompts allowed (e.g., for segmentation or detection models)?

No manual category-specific prompts are allowed.

Prompts must be:

Generated using a fixed template or automatic procedure
Possibly filled with automatically derived variables (e.g., category name)

Is it allowed to use different models for different categories (e.g., a multi-expert system)?

No.

Systems that route different categories to different models (even automatically) are not allowed.

Your method must use a single unified architecture, without category-specific expert selection, which includes the choice of pretrained backbone if applicable.

Is synthetic anomaly generation allowed?

Yes, under certain conditions.

Allowed:

Classical, class-agnostic methods (e.g., CutPaste, Perlin Noise)

Conditionally allowed:

More advanced methods (e.g., diffusion-based generation), provided that:
- They do not use knowledge of real anomalies from test data or the MVTec AD 2 paper.
- They follow a general, reproducible procedure

Can prompts or anomaly descriptions be generated automatically (e.g., using LLMs)?

Yes, if all of the following are satisfied:

Prompt generation is fully automatic
No manual editing or class-specific intervention
No use of test set information or prior knowledge of anomalies
The process is fully reproducible (including seeds, configs, etc.)

What should be included in the submission regarding prompts and generation methods?

You must include:

All prompt templates and generation logic
Any automatically generated prompts
Configuration details (e.g., seeds, parameters)

If results cannot be reproduced, the submission may be disqualified.

Can I use the test_public data to obtain a threshold in the regular setting?
No.
It is not admissible to use the test_public data for hyperparameter optimization such as threshold selection.

Can I use information/statistics about the test data during inference in the zero-shot setting?
No.
In the zero-shot setting, any use of information derived from previously observed target-category test samples during inference is not admissible. The model must not be updated during inference but remain in the initial zero-shot state.

Can I use validation images from MVTec AD 2 for threshold estimation in the zero-shot setting?
No.
It is not admissible to use any part of MVTec AD 2, including validation images, for any hyperparameter optimization in the zero-shot setting, which includes threshold selection.

VAND 4.0 Retail Track - Technical Guidelines

Technical Guidelines for the Retail Track are availbale for download via the following link:
https://drive.google.com/file/d/1It6U_fB-D7M5qQdZiasAOoZ6u-0DoCHT/view?usp=sharing

Please expand this section to see answers to Frequently Asked Questions about the Retail Track.

VAND 4.0 Retail Track - FAQs

Challenge Organizers

You can reach out to the Challenge Organizers via:

vand4-0-challenge@googlegroups.com

Please use this email for any questions around the challenge and restrain from contacting organizers individually since your request might not be answered in this case.