TAR

Official implementation of our CVPR 2026 paper "TAR: Token-Aware Refinement for Fine-grained Generalized Category Discovery"

📖 Introduction

This repository contains the official PyTorch implementation of TAR: Token-Aware Refinement for Fine-grained Generalized Category Discovery.

Our work focuses on fine-grained generalized category discovery and addresses attention artifacts that hinder models from capturing discriminative fine-grained information.

Key Contributions:

• 🚀 We reveal a previously overlooked challenge in Fine-grained Generalized Category Discovery, namely the attention artifact problem that hinders models from capturing discriminative fine-grained information.

• 🧠 We propose a plug-and-play method that can be easily integrated into existing models without modifying their architectures.

• 📊 Extensive experiments on multiple fine-grained benchmarks (e.g., CUB, FGVC-Aircraft, and Stanford Cars) demonstrate consistent and significant performance improvements.

📁 Project Structure

TAR/
│
├── clip/               # CLIP Model
├── scripts/            # training scripts
├── data/               # datasets
├── dataset_class_name/ # Generated data
├── util/               # functions
│
├── model.py            # Implementation of TAR
├── config.py           # Configuration file
├── requirements.txt
├── README.md
└── LICENSE

📂 Installation

git clone https://github.com/VectorYangYiStar/TAR.git
cd TAR

#remember to use Anaconda to create your virtual environment
pip install -r requirements.txt

🚀 Training

./scripts/run_aircraft.sh
./scripts/run_cifar10.sh
./scripts/run_cifar100.sh
./scripts/run_cub.sh
./scripts/run_herb.sh
./scripts/run_imagenet100.sh
./scripts/run_scars.sh

📄 Citation

If you find this project useful, please consider citing:

🤝 Acknowledgements

This project builds upon the following excellent works:

• GET
• SimGCD

📜 License

This project is released under the MIT License. See the LICENSE file for details.