Tipsomaly / README.md
Copilot
Copilot
Refer Tipsomaly checkpoint usage to GitHub
7ce1fc9
|
Raw History Blame Contribute Delete
3.26 kB
---
license: mit
library_name: pytorch
tags:
- anomaly-detection
- zero-shot
- prompt-learning
- mvtec
- visa
- arxiv:2602.03594
---
# Tipsomaly (ICASSP'26) learned prompt checkpoints
Learned prompt weights for the ICASSP'26 paper, **TIPS Over Tricks: Simple Prompts for Effective Zero-shot Anomaly Detection** (Tipsomaly), trained using the MVTec AD and VisA datasets. These are auxiliary prompt checkpoints for the official [Tipsomaly PyTorch implementation](https://github.com/Alireza99Salehi/Tipsomaly); they are not standalone model weights and do not include the TIPS vision-language backbone.
**Paper:** [Hugging Face paper page](https://huggingface.co/papers/2602.03594) · [arXiv](https://arxiv.org/abs/2602.03594) · [IEEE](https://ieeexplore.ieee.org/document/11461758)
**Demo:** [Tipsomaly on Hugging Face Spaces](https://huggingface.co/spaces/AlirezaSalehi99/Tipsomaly)
## Files
Each dataset directory contains training arguments and two learned-prompt checkpoints:
| Dataset | Files |
| --- | --- |
| MVTec AD | [`args.txt`](checkpoints/mvtec/args.txt), [`learnable_params_1.pth`](checkpoints/mvtec/learnable_params_1.pth), [`learnable_params_2.pth`](checkpoints/mvtec/learnable_params_2.pth) |
| VisA | [`args.txt`](checkpoints/visa/args.txt), [`learnable_params_1.pth`](checkpoints/visa/learnable_params_1.pth), [`learnable_params_2.pth`](checkpoints/visa/learnable_params_2.pth) |
`learnable_params_1.pth` and `learnable_params_2.pth` are saved at the end of training epochs 1 and 2, respectively. Each file stores the learned prompt tensor under the `learnable_prompts` key. The recorded training settings use 518-pixel images, the `concat` prompt-learning method, and segmentation loss.
## Download and use
Install and authenticate the Hugging Face CLI. From the Tipsomaly project directory, download the desired checkpoint files (replace `mvtec` with `visa` for VisA):
```bash
hf download AlirezaSalehi99/Tipsomaly \
checkpoints/mvtec/args.txt \
checkpoints/mvtec/learnable_params_2.pth \
--local-dir .
```
For instructions on preparing the environment and datasets and running either checkpoint, follow the [official GitHub repository's setup and inference instructions](https://github.com/Alireza99Salehi/Tipsomaly#-setup). The checkpoints contain learned prompts only; the TIPS backbone and evaluation datasets must be obtained separately.
## Training provenance
The checkpoint `args.txt` files record training with `prompt_learn_method=concat`, `cls_seg_los=seg`, seed 111, and image size 518. The MVTec and VisA weights were trained separately using their respective training datasets. See the paper and project repository for the method, full implementation, and experiment details.
## Citation
```bibtex
@article{salehi2026tips,
title={TIPS Over Tricks: Simple Prompts for Effective Zero-shot Anomaly Detection},
author={Salehi, Alireza and Karami, Ehsan and Noey, Sepehr and Noey, Sahand and Yamada, Makoto and Hosseini, Reshad and Sabokrou, Mohammad},
journal={arXiv preprint arXiv:2602.03594},
year={2026}
}
```
## License
The Tipsomaly project and these prompt checkpoints are released under the [MIT License](https://github.com/Alireza99Salehi/Tipsomaly/blob/main/LICENSE).