GFR-SAM: Training-Free Referring Camouflaged Object Segmentation via Cross-Image Prompting

Yilong Yang · Jianxin Tian · Shengchuan Zhang · Liujuan Cao

Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, Xiamen, China

🔥 News

• [2026/07/26] Paper and code are released.

⚙️ Environment Setup

Prerequisties

• Python 3.12 or higher
• PyTorch 2.7 or higher
• Transformers 5.0.0 or higher
• CUDA-compatible GPU with CUDA 12.6 or higher

1. Clone the repository and install the package:

git clone https://github.com/ynulonger/GFR-SAM.git
cd sam3/sam3
pip install -e .

2. Prepare checkpoints:

   GFR-SAM relies on frozen DINOv3 and SAM3 backbones. Please download the pretrained weights from the official repository:

   👉 https://huggingface.co/facebook/sam3

   👉 https://huggingface.co/facebook/dinov3-vitb16-pretrain-lvd1689m

   Place the weights and configuration files of the backbone under the hf-sam3 and hf-dinov3 folders.

3. Dataset preparation:

   Download the R2C7K dataset and replace the path of the dataset in the ref-sam3.ipynb.

4. Test and Visualisation:

   Run the ref-sam3.ipynb to perform in-context segmentation and visualise both intermediate and final output masks.

Citation

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

@inproceedings{cuttano2026insid3,
  title     = {GFR-SAM: Training-Free Referring Camouflaged Object Segmentation via Cross-Image Prompting},
  author    = {Yilong Yang, Jianxin Tian, Shengchuan Zhang, Liujuan Cao},
  booktitle = {34th ACM International Conference on Multimedia (ACM MM)},
  year      = {2026}
}

Acknowledgements

We gratefully acknowledge the contributions of the following open-source projects:

• DINOv3
• SAM3