RAISE: Realness Assessment for Image Synthesis and Evaluation

This repository contains the dataset and code for our paper RAISE: Realness Assessment for Image Synthesis and Evaluation, accepted at MIPR 2025.

Requirements

All dependencies are listed in requirements.txt. We recommend using Conda for environment management. Here's how to set it up:

# 1. Clone this repository
git clone https://github.com/annimukherjee/RAISE.git
cd RAISE

# 2. Create and activate a Conda environment
conda create -n raise python=3.10 -y
conda activate raise

# 3. Install required packages
pip install -r requirements.txt

RAISE Dataset

The RAISE dataset comprises 600 images out of which 480 are AI generated and 120 are real photographic images along with their corresponding subjective realness ratings as MOS scores. The images and corresponding ratings are made available under the /dataset directory.

There are 510 images in the training set and 90 images in the test set.

Image files:

• Real images: r1.png–r120.png
• AI-generated images: f1.png–f480.png

Dataset Structure

dataset/
├── images/
│   ├── test_images/
│   │   ├── f104.png
│   │   ├── ...
│   │   └── r83.png
│   └── train_images/
│       ├── f1.png
│       ├── ...
│       └── r115.png
└── ratings/
    ├── test.csv
    └── train.csv

Baseline Models

We provide the training as well as evaluation of the following baseline models for performing realness prediction on the RAISE dataset:

Model	Training Notebook	Testing Notebook
Decision Tree (handcrafted features)	02_ml-features-modelling.ipynb
CNN	0_base_cnn-train.ipynb	1_base_cnn-test.ipynb
Transfer Learning (ResNet-18)	0_transf-learning-train.ipynb	1_transf-learning-test.ipynb
JOINT Fine-Tuned (JOINT Rationality Branch)	0_joint-resnet-train.ipynb	1_joint-resnet-test.ipynb

A few important considerations for JOINT Fine-Tuned:

• Chen et al. (pdf) provide the weights to their JOINT Model (JOINT_2024.pth) in their official implementation. The weights can be found linked to in their README file. You can also directly download it from here (same link). Please place these weights in a new folder called ./models/04_NN-fine-tuned-JOINT/jointweights/.
• Also, please download the JOINT.py file from Chen et al.'s (pdf) official implementation and place this file in the ./models/04_NN-fine-tuned-JOINT/ directory.
• Our fine tuned weights for JOINT can be downloaded from this Google Drive Folder. Please place them in the ./models/04_NN-fine-tuned-JOINT/weights/ directory while testing the model.

Contact

For questions, please email the corresponding authors:

• Aniruddha Mukherjee: [email protected]
• Spriha Dubey: [email protected]

Citations

If you make use of the RAISE dataset or the code shared in this repository, please cite our paper as:

@misc{raise,
  title={RAISE: Realness Assessment for Image Synthesis and Evaluation},
  author={Aniruddha Mukherjee and Spriha Dubey and Somdyuti Paul},
  year={2025},
  eprint={2505.19233},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2505.19233},
}

License

This project is provided for non-commercial use only.

You may use, copy, modify, and share this project for personal, educational, or research purposes.
Commercial use of any part of this project is strictly prohibited without explicit written permission from the authors.

---

For questions about the dataset, models, or reproduction of results, please open an issue or contact the authors directly.