Request access to UniADet weights

Access is granted after you submit the form below. Please use an institutional email when possible.

UniADet MSDCS weights are released for research and evaluation. By requesting access you agree that: (1) you will cite the UniADet paper if you use these weights; (2) you will not redistribute the weights as a standalone package without this model card; (3) the authors may keep a record of this access request (Hugging Face username, name, affiliation, email, country, intended use, and request time) to measure academic and industrial adoption. This is an access-request log, not a per-file download log. Hugging Face also counts anonymous repository downloads automatically.

UniADet

Official MSDCS weights of One Language-Free Foundation Model Is Enough for Universal Vision Anomaly Detection, NeurIPS 2026.

This Hugging Face repository stores learnable MSDCS checkpoints only. Training and evaluation code is in a separate GitHub repository: gaobb/UniADet. Frozen CLIP / DINOv2 / DINOv3 backbones are not included here.

News

[2026-11-09]: Pre-trained MSDCS weights are released in this repository (gated: request access, then hf download).
[2026-09-25]: UniADet has been accepted to NeurIPS 2026. Official page: Poster.

Introduction

UniADet is a language-free universal (zero- and few-shot) visual anomaly detection framework. It outperforms language-dependent zero-/few-shot AD models while remaining simple and efficient: only 0.015M or 0.020M learnable parameters on top of a frozen foundation model.

Language-dependent vs language-free anomaly detection

Language prompts and text encoders are unnecessary. The same recipe works with CLIP, DINOv2-Reg, and DINOv3.
Global anomaly classification and local anomaly segmentation are fully decoupled across multi-scale hierarchical features (MSDCS: layer-wise cls/seg weights).
Few-shot UniADet is the first to outperform full-shot state-of-the-art on several industrial benchmarks.

Checkpoints in this repo

MSDCS uses 5 layers, dual cls/seg embeddings, and 2 tokens (normal / anomaly): (5 \times 2 \times 2 \times D). Files are fp32 .pth with a small pickle header. Each file stores {"msdcs": state_dict} only.

Backbone	(D)	Learnable params	File size	Frozen backbone (not in this repo)
CLIP `ViT-L/14@336px`	768	(1.5\times10^{-2}) M	64.5 KB	342.9 M
DINOv2-Reg `ViT-L-14`	1024	(2.0\times10^{-2}) M	84.5 KB	304.4 M
DINOv3 `ViT-L-16`	1024	(2.0\times10^{-2}) M	84.5 KB	303.2 M

Layout after download (same as $UNIADET_CKPT in the code repo):

train_on_<visa|mvtec>_zs_<backbone>_imagesize<H>_1015/epoch_15.pth

Path	Backbone	Trained on	Input size
`train_on_visa_zs_ViT-L-14@336px_imagesize518_1015/epoch_15.pth`	CLIP ViT-L/14@336px	VisA	518
`train_on_mvtec_zs_ViT-L-14@336px_imagesize518_1015/epoch_15.pth`	CLIP ViT-L/14@336px	MVTec	518
`train_on_visa_zs_ViT-L-14-DINOv2Reg_imagesize518_1015/epoch_15.pth`	DINOv2-Reg ViT-L/14	VisA	518
`train_on_mvtec_zs_ViT-L-14-DINOv2Reg_imagesize518_1015/epoch_15.pth`	DINOv2-Reg ViT-L/14	MVTec	518
`train_on_visa_zs_ViT-L-16-DINOv3_imagesize512_1015/epoch_15.pth`	DINOv3 ViT-L/16	VisA	512
`train_on_mvtec_zs_ViT-L-16-DINOv3_imagesize512_1015/epoch_15.pth`	DINOv3 ViT-L/16	MVTec	512

Paper protocol: train on VisA, test on the target set, unless you specifically need the MVTec-trained weights.

How to download

This repo is gated. Open https://huggingface.co/csgaobb/UniADet while logged in, fill the form, then:

pip install -U huggingface_hub
hf auth login
hf download csgaobb/UniADet --local-dir "$UNIADET_CKPT"

Please use hf download / snapshot_download (whole repo). That fetches config.json, which is the file Hugging Face uses to count downloads. Directly clicking a single .pth in the file browser may not increment the public download counter.

from huggingface_hub import snapshot_download

snapshot_download(repo_id="csgaobb/UniADet", local_dir="checkpoints")

Then follow the GitHub docs: environment, data, training / evaluation.

Results (Image-AUROC / Pixel-AUPR)

Backbone	Shot	MVTec-AD	VisA	Real-IAD
CLIP (ViT-L/14@336px)	0	92.4 / 42.8	88.0 / 28.0	78.6 / 33.6
CLIP (ViT-L/14@336px)	4	97.7 / 58.8	93.3 / 36.7	84.3 / 37.2
DINOv2 (Register ViT-L/14)	0	93.5 / 50.9	91.3 / 32.7	82.5 / 43.1
DINOv2 (Register ViT-L/14)	4	98.7 / 65.4	96.9 / 45.2	90.3 / 48.5
DINOv3 (ViT-L/16)	0	94.0 / 52.7	91.9 / 32.5	81.2 / 41.6
DINOv3 (ViT-L/16)	4	98.2 / 69.0	97.1 / 45.5	88.5 / 49.8

Full comparisons, ablations, and efficiency tables: GitHub README.

Intended use

Intended: academic research and reproduction of UniADet zero-shot / few-shot visual anomaly detection on public industrial and medical benchmarks.
Out of scope: clinical diagnosis, safety-critical deployment without additional validation, or treating these tiny MSDCS files as a standalone detector without the frozen backbone and official code.

Bias, risks, limitations

Weights were trained on auxiliary industrial data (VisA or MVTec). Cross-domain performance depends on the frozen backbone and the evaluation protocol in the paper.
Checkpoints contain MSDCS embeddings only. You still need the corresponding official CLIP / DINOv2-Reg / DINOv3 weights.
Gated access records who requested the repo, not every subsequent GET of a weight file.

Citation

If you find this work useful in your research, please consider citing:

@inproceedings{uniadet,
  title={One Language-Free Foundation Model Is Enough for Universal Vision Anomaly Detection},
  author={Gao, Bin-Bin},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
  year={2026},
  url={https://neurips.cc/virtual/2026/poster/149035}
}

Downloads last month: 8

csgaobb
/

UniADet