Official PyTorch implementation of "Towards Explainable Image Aesthetics Assessment with Attribute-oriented Critiques Generation" (IEEE TCSVT 2025).

• 🎉 We release the multi-attribute aesthetic critiques generation model with pre-trained weights and training data!
• 🎉 Our CG-IAA paper was accepted by IEEE TCSVT!
• [Coming Soon] The complete aesthetic assessment model will be released soon.

CG-IAA addresses a critical challenge in image aesthetics assessment: How can we leverage the power of multimodal learning when aesthetic critiques are unavailable? Our solution generates high-quality aesthetic critiques from multiple attribute perspectives, enabling both accurate aesthetic prediction and enhanced model explainability.

• Multi-Attribute Aesthetic Critiques Generation: We propose a CLIP-based model that generates diverse aesthetic critiques from four different perspectives:

  • 🎨 Color and Light: Color harmony, saturation, lighting quality
  • 📐 Composition: Layout, balance, structural elements
  • 🔍 Depth and Focus: Depth of field, focus, blur effects
  • ⭐ General Feelings: Overall aesthetic impression and quality

• Enhanced Explainability: Generated critiques provide human-readable explanations for aesthetic judgments, making the model more transparent and interpretable.

The CG-IAA framework consists of three main components:

1. VLAP (Vision-Language Aesthetic Pretraining): Fine-tune CLIP on aesthetic data
2. MAEL (Multi-Attribute Experts Learning): Train attribute-specific expert models
3. MAP (Multimodal Aesthetics Prediction): Fuse visual and textual features for final prediction

1. Aesthetic Critiques Generation Model - Multi-attribute aesthetic critiques generation

   • Pre-trained model weights
   • Inference code for single image processing

2. Training Data - Large-scale multi-attribute aesthetic critique dataset

   • ~150K critiques for Color and Light
   • ~100K critiques for Composition
   • ~120K critiques for Depth and Focus
   • ~570K critiques for General Feelings
   • Total: ~940K aesthetic critiques with attribute annotations

• Complete aesthetic assessment model

# Clone the repository
git clone https://github.com/your-username/CG-IAA.git
cd CG-IAA

# Create and activate conda environment
conda env create -f environment.yml
conda activate cg-iaa

Download the pre-trained model weights from Google Drive and place them in the checkpoints/ directory:

📥 Download Model Weights The checkpoints directory should contain:

checkpoints/
├── base_model.pt          # Base model
├── color.pt              # Color expert model
├── composition.pt        # Composition expert model
├── dof.pt               # Depth of Field expert model
└── general.pt           # General expert model

If you want to train your own models, download our multi-attribute aesthetic critique dataset:

📥 Download Training Data

Generate aesthetic critiques for a single image:

python caption_inference.py --image_path samples/1.jpg

Output:

================================================================================
Multi-Attribute Aesthetic Captions for: samples/1.jpg
================================================================================

[Color]

[Composition]

[Depth of Field]

[General]

================================================================================

Our generated aesthetic critiques achieve competitive performance when used alone for IAA task:

Tested on AVA database using text-only input

Our released multi-attribute aesthetic critique dataset is organized as follows:

data/
├── color.json           # Color and Light critiques
├── composition.json     # Composition critiques
├── dof.json            # Depth and Focus critiques
└── general.json        # General Feelings critiques

Each JSON file contains entries in the following format:

[
  {
    "id": 0,
    "img_id": "773931",
    "caption": "Image feels a tad dark, which I dont think helps this image for me."
  },
  ...
]

CG-IAA is built upon the following excellent open-source projects:

• CLIP - Contrastive Language-Image Pre-training
• ClipCap - CLIP Prefix for Image Captioning
• timm - PyTorch Image Models

If you find our work useful, please consider citing our paper:

@article{li2025cgiaa,
  author={Li, Leida and Sheng, Xiangfei and Chen, Pengfei and Wu, Jinjian and Dong, Weisheng},
  journal={IEEE Transactions on Circuits and Systems for Video Technology}, 
  title={Towards Explainable Image Aesthetics Assessment With Attribute-Oriented Critiques Generation}, 
  year={2025},
  volume={35},
  number={2},
  pages={1464-1477}
}