Spaces:

---

jiang-cc

/

AD-Copilot

like 1

Running on Zero

AD-Copilot is an open-source vision-language assistant for industrial anomaly detection (IAD), presented in our paper "AD-Copilot: A Vision-Language Assistant for Industrial Anomaly Detection via Visual In-context Comparison" (arXiv:2603.13779, under review at IEEE Transactions on Image Processing).

What it does: Given a normal reference image and a query image of an industrial product, AD-Copilot detects, classifies, describes, and localizes anomalies through natural language interaction — functioning as an AI-powered industrial inspector.

Key contributions:

• A novel Comparison Encoder using cross-attention between paired image features, enabling fine-grained visual comparison at the encoding stage rather than only in the language space
• Chat-AD: a large-scale multimodal IAD dataset with 620k+ samples across 327 industrial categories - MMAD-BBox: a new benchmark for fine-grained anomaly localization with bounding-box evaluation
• Achieves 82.3% accuracy on the MMAD benchmark, surpassing all proprietary and open-source models (including GPT-4o, Gemini 1.5 Pro) and exceeding ordinary human performance, without any data leakage
• 3.35× improvement over baseline on anomaly localization (MMAD-BBox)

Why GPU is needed: The model (7B, ~16GB bfloat16) requires GPU for inference. We are building an interactive Gradio demo on HuggingFace Spaces (jiang-cc/AD-Copilot) so researchers and industrial practitioners can try AD-Copilot directly in the browser.

Open resources:

• Paper: https://arxiv.org/abs/2603.13779
• Model: https://huggingface.co/jiang-cc/AD-Copilot
• Code & data: https://github.com/jam-cc/AD-Copilot

This is an academic project from Southern University of Science and Technology (SUSTech). We would greatly appreciate ZeroGPU access to make our demo publicly available for the community.