yzc002/PRAC

Personalized Image Aesthetic Assessment via Preference-rich Sample Mining and Cohort Merging

Zhichao Yang^1†, Tianjiao Gu^1†, Zhixianhe Zhang¹, Xiangfei Sheng¹, Pengfei Chen¹, Leida Li^1,2*

¹School of Artificial Intelligence, ²State Key Laboratory of EMIM, Xidian University

^†Equal contribution    ^*Corresponding author

News

• [2026-10-03] ✨✨ The Code and Pre-trained Weights, are now publicly available.
• [2026-08-28] 🎉🎉 Congratulations! Our paper has been accepted for an Oral Presentation at ACM MM 2026.
• [2026-07-10] 🎉🎉 Our paper, "Personalized Image Aesthetic Assessment via Preference-rich Sample Mining and Cohort Merging", has been accepted to ACMMM 2026!

Introduction

This release contains the PARA implementation of PRAC: the Generic Aesthetic Predictor (GIAA), PreferSelect, PreferMerge, data loaders, and experiment configurations. Run commands from the repository root.

Quick Start

1. Create the environment

The following environment matches the released configuration. Install the PyTorch build that matches the CUDA runtime available to the target GPU; the example uses PyTorch 2.4.1 with CUDA 12.1.

conda create -n prac python=3.10 -y
conda activate prac
conda install pytorch==2.4.1 torchvision==0.19.1 pytorch-cuda=12.1 -c pytorch -c nvidia
pip install -r requirements.txt

2. Download the model and data

Download the mPLUG-Owl3-7B-241101 base checkpoint and place it at models/mPLUG-Owl3-7B, or set model.model_path in the YAML files to another compatible local path. Then you can download the pretrained adapter from Baidu Netdisk (pwd=3rd3).

Obtain the PARA dataset from this link and extract the archive's directory to data. The loaders also need derived files. Generate them from the original annotations with:

python scripts/prepare_giaa_labels.py --source data/PARA --output data/PARA

The final data layout is:

data/PARA/
├── README.md                                # original dataset README
├── imgs/<session_id>/<image_name>           # original images
├── annotation/                              # original PARA annotations
├── para_giaa_train_discretized.csv          # from annotation/PARA-GiaaTrain.csv
├── para_giaa_test_discretized.csv           # from annotation/PARA-GiaaTest.csv
└── ...                                      # other original PARA files

The personalized-data loaders read the original annotation/PARA-Images.csv directly. User prompts are kept in the repository at data/user_info_para.json; the label-generation script does not copy or modify either file.

3. Run the inference demo

After the base checkpoint and PARA images are available, the demo automatically scores the first image below data/PARA/imgs:

python demo_inference.py

To use a particular image, a trained adapter, or a generated rationale:

python demo_inference.py \
  --image data/PARA/imgs/<session_id>/<image_name> \
  --adapter-path runs/para/giaa_adapter \
  --explain

Repository Structure

.
├── configs/
│   ├── giaa_train_config/       GIAA training configuration
│   ├── giaa_eval_config/        GIAA evaluation and demo configuration
│   ├── preferselect_config/     CCM/PDM data and model configuration
│   └── prefermerge_config/      LoRA-pool, FIM, and merge configuration
├── data/                        Dataset layout and download notes
├── data_loader/                 GIAA and personalized PARA data loaders
├── modules/MplugOwl3ForGIAA.py   mPLUG-Owl3 wrapper and LoRA utilities
├── demo_inference.py            Single-image distribution and score demo
├── mplug3_train_giaaBackbone.py Train the generic aesthetic predictor
├── mplug3_eval_giaaBackbone.py  Evaluate a GIAA adapter
├── preferSelect_CCM.py          Compute collective controversy (CCM)
├── preferSelect_PDM_fast.py     Compute cached personalized deviation (PDM)
├── preferSelect_weights_ablation.py
│                                Combine CCM and PDM for all alpha values
├── preferSelect.py              Export ranked preference-rich samples
├── preferMerge.py               Train LoRA pool or run adapter merging
├── preferMerge_FIM_fixed.py     Compute FIM preference similarities
├── preferMerge_choice_fast.py   Select a relevance/diversity cohort
├── utils.py                     Shared losses, metrics, and configuration helpers
├── CITATION.cff                 Citation metadata
├── LICENSE                      MIT license for the PRAC source code
└── docs/                        Project page and reproduction guide

Re-Training

1. Generic Aesthetic Predictor

Edit configs/giaa_train_config/para.yaml if the model, data, batch size, or output path differs from the defaults, then run:

python mplug3_train_giaaBackbone.py

The trainer writes LoRA checkpoints below runs/para/giaa. Evaluate a selected adapter with:

python mplug3_eval_giaaBackbone.py

2. Preference-rich Sample Mining (PreferSelect)

Run the commands below:

python preferSelect_CCM.py
python preferSelect_PDM_fast.py
python preferSelect_weights_ablation.py
python preferSelect.py

3. Aesthetically-resonant Cohort Merging (PreferMerge)

Run the commands below:

python preferMerge_FIM_fixed.py
python preferMerge_choice_fast.py --top-n 6 --beta 0.5
python preferMerge.py

Citation

If this code or paper helps your research, please cite:

@inproceedings{yang2026personalized,
  title={Personalized Image Aesthetic Assessment via Preference-rich Sample Mining and Cohort Merging},
  author={Yang, Zhichao and Gu, Tianjiao and Zhang, Zhixianhe and Sheng, Xiangfei and Chen, Pengfei and Li, Leida},
  booktitle={Proceedings of the 34th ACM International Conference on Multimedia},
  year={2026}
}

Citation metadata is also available in CITATION.cff.

License

The PRAC source code is available under the MIT License.