MG-VLQA: Multi-Granularity Quality Assessment for Image Compression via Visual Language Models

An image quality assessment framework leveraging visual-language models for multi-granularity semantic fidelity.

Hanfei Li¹  ·  Anle Ke¹  ·  Jiawen Gu²  ·  Chao Zhou²  ·  Tong Chen¹  ·  Zhan Ma¹  · 

¹ Nanjing University     ² Kuaishou Technology

🚀Quick Start

📁Prerequisites

Before installing the dependencies, please prepare the model weights.

Create the following directory structure in the project root:

weights/
├── Gemma/
└── SAM/

Download the required model weights from Hugging Face and place them in the corresponding folders:

Gemma weights: https://huggingface.co/google/gemma-2b-it

SAM weights: https://huggingface.co/google-bert/bert-base-uncased

🧪Environment Requirements

Make sure your environment meets the following requirements:

• Python >= 3.9
• PyTorch == 2.6.0
• transformers == 4.53.0
• timm == 1.0.6
• six
• accelerate

Then, navigate to the GroundingDINO directory and install it in editable mode:

cd GroundingDINO
pip install -e .