FLAME: Frozen Large Language Models Enable Data-Efficient Language-Image Pre-training
Anjia Cao, Xing Wei, Zhiheng Ma

• [2025/03/09] Release training codes and scripts.
• [2025/03/08] Recaptioned CC3M and Recaptioned YFCC15M on Hugging Face.
• [2025/02/27] Accepted by CVPR 2025.
• [2024/11/28] Model on Hugging Face.
• [2024/11/28] Release evaluation codes and scripts.
• [2024/11/18] Paper on arXiv.

• 🔥 Leveraging frozen LLMs to naturally process long text inputs.
• 🔥 Generalizing from monolingual training to multilingual evaluation.
• 🔥 Strong improvement on long/short-context image-text retrieval, image classification, and multilingual scenarios.

• Release training code and data.
• Release evaluation code.
• Release pre-trained checkpoints.

git clone https://github.com/MIV-XJTU/FLAME.git
cd FLAME
conda create -n flame python=3.10 -y
conda activate flame
make install
make install-training
make install-test

See Training.md.

See Evaluation.md.

The project is under a standard Creative Common CC-BY-4.0 License.

If you find our work helpful for your research, please consider giving a star and citation.

@inproceedings{cao2025flame,
  title={FLAME: Frozen Large Language Models Enable Data-Efficient Language-Image Pre-training},
  author={Cao, Anjia and Wei, Xing and Ma, Zhiheng},
  booktitle={CVPR},
  year={2025}
}

This project is based on open_clip, and thanks for the nice work! We also thank CLIP_benchmark, DreamLIP, Long-CLIP, PromptEOL, and MiniCPM-V for their codes.