LLaVA-Mini: Efficient Image and Video Large Multimodal Models with One Vision Token

Shaolei Zhang, Qingkai Fang, Zhe Yang, Yang Feng*

LLaVA-Mini is a unified large multimodal model that can support the understanding of images, high-resolution images, and videos in an efficient manner. Guided by the interpretability within LMM, LLaVA-Mini significantly improves efficiency while ensuring vision capabilities. Model and demo of LLaVA-Mini are available now!

[!Note] LLaVA-Mini only requires 1 token to represent each image, which improves the efficiency of image and video understanding, including:

• Computational effort: 77% FLOPs reduction
• Response latency: reduce from 100 milliseconds to 40 milliseconds
• VRAM memory usage: reduce from 360 MB/image to 0.6 MB/image, support 3-hour video processing

💡Highlight:

1. Good Performance: LLaVA-Mini achieves performance comparable to LLaVA-v1.5 while using only 1 vision token instead of 576 (compression rate of 0.17%).
2. High Efficiency: LLaVA-Mini can reduce FLOPs by 77%, deliver low-latency responses within 40 milliseconds, and process over 10,000 frames of video on the GPU hardware with 24GB of memory.
3. Insights: To develop LLaVA-Mini, which reduces vision tokens while maintaining visual understanding, we conduct a preliminary analysis to explore how large multimodal models (LMMs) process visual tokens. Please refer to our paper for a detailed analysis and our conclusions.

🖥 Demo

• Download LLaVA-Mini model from here.

• Run these scripts and Interact with LLaVA-Mini in your browser:

  # Launch a controller
  python -m llavamini.serve.controller --host 0.0.0.0 --port 10000 &
  
  # Build the API of LLaVA-Mini, if the VRAM memory is less than 20GB, try using --load-8bit
  CUDA_VISIBLE_DEVICES=0  python -m llavamini.serve.model_worker --host 0.0.0.0 --controller http://localhost:10000 --port 40000 --worker http://localhost:40000 --model-path ICTNLP/llava-mini-llama-3.1-8b --model-name llava-mini &
  
  # Start the interactive interface
  python -m llavamini.serve.gradio_web_server --controller http://localhost:10000 --model-list-mode reload  --port 7860
  

🔥 Quick Start

Requirements

• Install packages:

  conda create -n llavamini python=3.10 -y
  conda activate llavamini
  pip install -e .
  pip install -e ".[train]"
  pip install flash-attn --no-build-isolation
  

Command Interaction

• Image understanding, using --image-file.

• If the VRAM memory is less than 20GB, try using --load-8bit.

  # Image Understanding
  CUDA_VISIBLE_DEVICES=0 python llavamini/eval/run_llava_mini.py \
      --model-path  ICTNLP/llava-mini-llama-3.1-8b \
      --image-file llavamini/serve/examples/baby_cake.png \
      --conv-mode llava_llama_3_1 --model-name "llava-mini" \
      --query "What's the text on the cake?"
  

• Video understanding, using --video-file:

  # Video Understanding
  CUDA_VISIBLE_DEVICES=0 python llavamini/eval/run_llava_mini.py \
      --model-path  ICTNLP/llava-mini-llama-3.1-8b \
      --video-file llavamini/serve/examples/fifa.mp4 \
      --conv-mode llava_llama_3_1 --model-name "llava-mini" \
      --query "What happened in this video?"
  

Reproduction and Evaluation

• Refer to Evaluation.md for the evaluation of LLaVA-Mini on image/video benchmarks.

Cases

• LLaVA-Mini achieves high-quality image understanding and video understanding.

More cases

• LLaVA-Mini dynamically compresses image to capture important visual information (brighter areas are more heavily weighted during compression).

🤝 Acknowledgement

• LLaVA: LLaVA-Mini is built upon LLaVA codebase, a large language and vision assistant.
• Video-ChatGPT: The training of LLaVA-Mini involves the video instruction data provided by Video-ChatGPT.
• LLaVA-OneVision: The training of LLaVA-Mini involves the image instruction data provided by LLaVA-OneVision.

🖋Citation

If this repository is useful for you, please cite as:

@misc{llavamini,
      title={LLaVA-Mini: Efficient Image and Video Large Multimodal Models with One Vision Token}, 
      author={Shaolei Zhang and Qingkai Fang and Zhe Yang and Yang Feng},
      year={2025},
      eprint={2501.03895},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2501.03895}, 
}

If you have any questions, please feel free to submit an issue or contact zhangshaolei20z@ict.ac.cn.