---
license: cc-by-nc-sa-4.0
tags:
  - anomaly-detection
  - industrial-inspection
  - mvtec-ad
  - mvtec-loco
  - patchcore
  - csad
---

# AnoDe model bundle

The models, thresholds and sample images served by the AnoDe web platform
(Space: [peanut2butter9/anode](https://huggingface.co/spaces/peanut2butter9/anode)).

| Products | Model | Files | Decision threshold |
|---|---|---|---|
| 15 MVTec AD categories | PatchCore (our implementation; WideResNet-50, 10% coreset) | `models/patchcore/<category>.pt` (fp16 memory bank) | mean + 3 std of scores on 10% held-out normal training images |
| 5 MVTec LOCO AD categories | CSAD, released checkpoints (Hsieh & Lai, BMVC 2024) | `models/csad/<category>.pth` | mean + 3 std of scores on the official validation split |

`manifest.json` lists every product with its model file, threshold, the
validation scores behind it, and its sample images (`samples/`, 512 px PNG,
with ground-truth masks for anomalous images).

The benchmark behind these choices and the reasons GLASS is not deployed
(its released checkpoints leave no normal images to calibrate a threshold)
are documented in the project repository.

## Licenses

- Sample images, and models trained on them: MVTec AD (Bergmann et al., IJCV
  2021) and MVTec LOCO AD (Bergmann et al., IJCV 2022), (c) MVTec Software
  GmbH, [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).
  Non-commercial use only.
- CSAD checkpoints: from https://github.com/Tokichan/CSAD, Apache License 2.0
  (`models/csad/LICENSE`).