🔥❄️🌡️ ICTA2Net

Thinking Aesthetics Assessment of Image Color Temperature

Models, Datasets, Benchmarks

---

📢 Announcement | 项目公告

🎉 The full paper and official code have been released in this repository.
🎉 论文全文与官方代码已正式发布，欢迎查阅与使用！

---

📁 Resource Links | 项目资源

🔖 Type	🔗 Access Link
📄 Paper / 论文PDF	Download PDF
🌐 Project Page / 项目主页	ICTA2Net.github.io
🧠 Pre-trained Weights / 预训练模型权重	Google Drive, 百度网盘
📊 Dataset / 数据集	百度网盘

---

🖼️ Visualization Gallery

---

Dataset Overview

Figure 1. Dataset Overview. Our dataset consists of multiple sets of images with different white balance shifts, along with their corresponding high-quality aesthetic reference images. The t-SNE visualization of the images at various color temperatures in the dataset is shown in the figure. This dataset is constructed from linear raw RGB images in the MIT-Adobe FiveK and PPR10K datasets. By precisely simulating the camera ISP process, we generate multiple rendered versions of each image with varying color temperatures.

---

Model Architecture

Figure 2. Overall framework of ICTA2Net, comprising four components: a Color Temperature Encoder for capturing color temperature variations; a Contextual Awareness Module (including Visual Encoder, Text Encoder, and Text Denoise Model); a Cross-Modal Fusion Module for visual–textual integration; and a Pairwise Ranking Predictor for aesthetic preference estimation.

---

Aesthetic Ranking Results

Figure 3. Visualization of model ranking results: aesthetic scores decrease progressively from left to right and top to bottom.

---

🚀 Quick Start

1. Environment Preparation

Recommended: Python 3.9+, PyTorch 1.12+, CUDA 11.6+

# Clone repository
git clone https://github.com/chasecjg/ICTA2Net.git
cd ICTA2Net

2. Dataset Setup

1. 📥 Download: Get the dataset from the provided link.

2. 📂 Unzip: Extract to the specified directory (update dataset_root in options.py).

3. 📊 Training Splits: Two splits are available:

   • 📜 train_42.csv: Full dataset (42k samples)
   • ⚡ train_8.csv: Optimized subset (8k samples, recommended for quick training)

3. Model Training

# Modify hyperparameters in options.py (e.g., resume, weight path)
python train.py

4. Inference & Evaluation

# Adjust test parameters in options.py (e.g., test dataset path)
python test.py

📝 Citation

@inproceedings{cheng2026thinking,
  title     = {Thinking Aesthetics Assessment of Image Color Temperature: Models, Datasets and Benchmarks},
  author    = {Cheng, Jinguang and Li, Chunxiao and He, Shuai and Chen, Taiyu and Ming, Anlong},
  booktitle = {Proceedings of the 40th AAAI Conference on Artificial Intelligence (AAAI)},
  year      = {2026},
  note      = {Poster},
  url       = {https://github.com/chasecjg/ICTA2Net/tree/main}
}

---

📫 Welcome to star, fork and collaborate!

🔝 Back to Top