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

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

[!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

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:
  bash

  1# Launch a controller 2python -m llavamini.serve.controller --host 0.0.0.0 --port 10000 & 3 4# Build the API of LLaVA-Mini 5CUDA_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 & 6 7# Start the interactive interface 8python -m llavamini.serve.gradio_web_server --controller http://localhost:10000 --model-list-mode reload --port 7860

🔥 Quick Start

Requirements

• Install packages:
  bash

  1conda create -n llavamini python=3.10 -y 2conda activate llavamini 3pip install -e . 4pip install -e ".[train]" 5pip install flash-attn --no-build-isolation

Command Interaction

• Image understanding, using --image-file :
  bash

  1# Image Understanding 2CUDA_VISIBLE_DEVICES=0 python llavamini/eval/run_llava_mini.py \ 3 --model-path ICTNLP/llava-mini-llama-3.1-8b \ 4 --image-file llavamini/serve/examples/baby_cake.png \ 5 --conv-mode llava_llama_3_1 --model-name "llava-mini" \ 6 --query "What's the text on the cake?"
• Video understanding, using --video-file :
  bash

  1# Video Understanding 2CUDA_VISIBLE_DEVICES=0 python llavamini/eval/run_llava_mini.py \ 3 --model-path ICTNLP/llava-mini-llama-3.1-8b \ 4 --video-file llavamini/serve/examples/fifa.mp4 \ 5 --conv-mode llava_llama_3_1 --model-name "llava-mini" \ 6 --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).

🖋Citation

@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}, 
}