Coherence Assessment of Generated Realistic Images

This repository contains a dataset of 4,368 PNG images generated based on the COCO dataset, created to assess the coherence and realism of AI-generated imagery. The dataset is freely accessible and available for anyone to use and explore.

This dataset accompanies the article titled "Easy for us, complex for AI: Assessing the coherence of generated realistic images", which surveys the challenges of evaluating realism in AI-generated images using current metrics and criteria.

Article Abstract

The existence of several effective AI methods that generate realistic images has garnered significant attention in recent years. Notable generative approaches utilizing different AI techniques, including Variational Autoencoders (VAEs), Generative Adversarial Networks (GANs), style-based generators, transformers, and diffusion models, have become popular for producing high-quality and diverse images. However, while humans can naturally assess the realistic qualities of images, this task remains challenging for quantitative automatic systems. This document presents a survey of current metrics and criteria for assessing the coherence of AI-generated images in terms of realism and quality. This review highlights how the metrics for realism in generated images fall short of understanding the high coherence that is essential for creating a realistic and immersive visual experience indistinguishable from real-world scenes. This follows Moravec’s paradox, which states that tasks easy for humans, such as pattern recognition, are often difficult for computers, which proves that the search for realism metrics keeps going. Specifically, this review discusses how the coherence of generated realistic images contributes to their believability by human perception, showing various levels of structural consistency, harmony, and logical connections between elements such as texture, color, lighting, and temporal aspects to create a cohesive scene that aligns with our understanding of the real world.

Dataset Overview

• Source: Images generated using the COCO dataset as a reference.
• Total Images: 4,368
• Format: PNG
• Purpose: The dataset can be used for tasks such as:
  • Research on AI image generation
  • Realism and coherence assessment in AI-generated images
  • Developing or evaluating image quality metrics
  • Training machine learning models
  • Studies on visual coherence and realism

Dataset Structure

The dataset is organized as follows:

/dataset
  /Generated\ Images
    1.png
    2.png
    ...
    4368.png

Usage

You are free to use, modify, and redistribute the dataset for any project. We encourage researchers, developers, and the broader AI community to explore this dataset for tasks related to image quality assessment, coherence analysis, and AI-generated imagery evaluation.

How to Use

1. Clone the repository:

   git clone https://github.com/PocketNugget/Coherence-assessment-of-generated-realistic-images.git

2. Access the images in the dataset/images/ directory.

License

This dataset is released under the MIT License, allowing you to freely use, modify, and distribute the images as long as proper credit is given.

Contributions

Contributions to the dataset or related research are welcome! If you have any suggestions or improvements, feel free to open an issue or submit a pull request.

Contact

For any questions or inquiries, please contact Goben Diego Constantino Aguirre at goben.ca.pkn@hotmail.com.

---