Res²CLIP: Few-Shot Generalist Anomaly Detection with Residual-to-Residual Alignment

Res²CLIP is a few-shot generalist anomaly detection framework that aligns visual residuals with text residuals using a frozen CLIP backbone. It supports two modes:

Mode	Symbol	Description
training-free	Res2CLIP*	Direct three-branch fusion on frozen CLIP features without fine-tuning.
finetune	Res2CLIP†	Lightweight adapters trained on an auxiliary dataset for higher performance.

---

Environment Preparation

conda create -n res2clip python=3.10
conda activate res2clip
pip install -r requirements.txt

---

Dataset Preparation

Dataset metadata JSON files are generated following the same procedure as AnomalyCLIP, please refer to AnomalyCLIP for scripts and instructions.

---

Backbone Preparation

We use the CLIP ViT-L/14@336px backbone. The model is downloaded automatically on first run to ./clip_model/ViT-L-14-336px.pt (or download manually from the OpenAI CLIP releases and place it there).

---

Training

Edit paths in train.sh, then:

bash train.sh

Adapters are trained separately on MVTec AD and VisA. Checkpoints are saved to ./checkpoints/{mvtec,visa}/.

---

Evaluation

Training-free (Res²CLIP^*):

bash test_trainingfree.sh

Fine-tuned (Res²CLIP^†):

bash test_finetune.sh

---

Acknowledgement

We thank AnomalyCLIP for their open-source codebase, on which clip_lib/ is based.

---

Citation

If you think this work is helpful to you, please consider citing our paper.

@article{liu2026res2clip,
  title={Res$^2$CLIP: Few-Shot Generalist Anomaly Detection with Residual-to-Residual Alignment},
  author={Liu, Xinyue and Wang, Jianyuan and Leng, Biao and Zhang, Shuo},
  journal={arXiv preprint arXiv:2605.16171},
  year={2026}
}