Crane and Crane+ learned prompt checkpoints (BMVC'26)

Checkpoints for Crane: Context-Guided Prompt Learning and Attention Refinement for Zero-Shot Anomaly Detection, a zero-shot anomaly detection method built on CLIP. The paper was published at BMVC 2026. Crane+ adds DINOv2-based attention refinement. These are learned prompt checkpoints for the official Crane implementation, not standalone model weights; the CLIP and optional DINOv2 backbones are obtained separately by the implementation.

Paper: Hugging Face paper page · arXiv · Project repository

Demo: Crane on Hugging Face Spaces

Checkpoints

The repository contains checkpoints saved after epochs 1–5 for each training configuration:

Directory	Variant	Training dataset
trained_on_mvtec_crane/	Crane	MVTec AD
trained_on_visa_crane/	Crane	VisA
trained_on_mvtec_cranep/	Crane+ (with DINOv2)	MVTec AD
trained_on_visa_cranep/	Crane+ (with DINOv2)	VisA

Each directory contains epoch_1.pth through epoch_5.pth. The evaluation code loads the selected file as ./checkpoints/<model_name>/epoch_<epoch>.pth.

Download

To download all five checkpoints for one configuration into the corresponding project directory, install and authenticate the Hugging Face CLI, then run (replace the directory name to select a different dataset or variant):

hf download AlirezaSalehi99/Crane \
  --include 'checkpoints/trained_on_mvtec_crane/*' \
  --local-dir .

For instructions on configuring the environment and running any Crane or Crane+ checkpoint, follow the inference instructions in the official GitHub repository.

License

The project and checkpoints are released under the MIT License.