This repository is the official PyTorch implementation of MA-AGIQA:Large Multi-modality Model Assisted AI-Generated Image Quality Assessment.
Our experiments are based on two conda environment. One for Semantic Feature Extraction and another for Train and Test.
git clone https://github.com/Q-Future/Q-Align.git
cd Q-Align
pip install -e .
git clone https://github.com/wangpuyi/MA-AGIQA.git
cd MA-AGIQA
pip install -r requirements.txt
We use official mPLUG-Owl2 to extract semantic features. The feature extraction codes are based on Q-Align, great thanks to them!
First, download and transfer root to Q-Align (You should download their repository as said in Requirements.)
cd Q-Align
then
- put json files containing information of training data to
Q-Align/playground/data/test_jsons
likeQ-Align/playground/data/test_jsons/AGIQA_3k.json
. - put
getFeature.py
likeQ-Align/q_align/evaluate/getFeature.py
.
You can find them under q_align
file in this repository and get semantic feature by
python "q_align/evaluate/getFeature.py"
if you have error when connect to Hugging Face, we recommand you use
HF_ENDPOINT=https://hf-mirror.com python "q_align/evaluate/getFeature.py"
Download and transfer root to MA-AGIQA. If you've download this repository, just implement the "cd" code.
cd MA-AGIQA
Train and Test
python train.py
- release the checkpoints
- simplify codes for friendly usage
If you find our code or model useful for your research, please cite:
@misc{wang2024large,
title={Large Multi-modality Model Assisted AI-Generated Image Quality Assessment},
author={Puyi Wang and Wei Sun and Zicheng Zhang and Jun Jia and Yanwei Jiang and Zhichao Zhang and Xiongkuo Min and Guangtao Zhai},
year={2024},
eprint={2404.17762},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
Part of our code are based on MANIQA and Q-Align. Thanks for their awesome work!