README.md

December 30, 2025 ยท View on GitHub

Bridging Cognitive Gap: Hierarchical Description Learning for Artistic Image Aesthetics Assessment

Paper Code Model Dataset

Motivation

Method

๐Ÿ”ง Installation

If you only need to infer / evaluate:

pip install -e .

For training, you need to further install additional dependencies as follows:

pip install -e ".[train]"
pip install flash_attn --no-build-isolation

Training, Inference & Evaluation

๐Ÿ’ช Preprocess

The APDD dataset has been processed APDD/apdd_score.json. If you use the Score-Based Distribution Estimation algorithm (ยง3.2) for your dataset, you can use the example script.

๐Ÿ”ฅ Training

Fine-tuning needs to download the mPLUG-Owl2 weights as in Pretrained Weights. Put it under the ModelZoo folder.

  • Only 2 RTX3090 GPUs are required.
sh train.sh

๐Ÿš€ Inference

After training, you can infer with:

sh infer.sh $ONE_GPU_ID

๐Ÿ” Evaluation

After inference, you can evaluate SRCC / PLCC for quality score.

sh eval.sh