# Exploring the Naturalness of AI-Generated Images

**Authors:** Zijian Chen, Wei Sun, Haoning Wu, Zicheng Zhang, Jun Jia, Xiongkuo Min, Guangtao Zhai, Wenjun Zhang  
**arXiv:** [2312.05476](https://arxiv.org/abs/2312.05476)  
**Date:** 2023-12-09  
**Summary page:** https://www.summarizepaper.com/en/paper/2312.05476v1/

## Key points

- Authors address the challenge of assessing the naturalness of Artificial Intelligence-Generated Images (AGIs)
- Constructed AI-Generated Image Naturalness (AGIN) database through a large-scale subjective study
- AGIN database verifies that naturalness is affected by both types of distortions
- Proposed Joint Objective Image Naturalness evaluaTor (JOINT) to learn naturalness aligning with human ratings
- JOINT outperforms baselines in providing more subjectively consistent results on naturalness assessment
- Appendix section provides additional details for construction of AGIN database
- Collected multi-sourced images from five tasks to cover diverse appearances of naturalness issues
- Chose 20 keywords and created 10 prompts for each keyword using GPT3.5 for text-to-image models
- Generated prompts were manually examined for compliance before collecting AI-generated images
- Proposed JOINT++ as an extension to JOINT, incorporating additional features into their model
- Paper contributes to understanding and evaluating the naturalness of AGIs by presenting a benchmark database and objective evaluators aligned with human opinions

## Summary

In this paper, the authors address the challenge of assessing the naturalness of Artificial Intelligence-Generated Images (AGIs). They construct the AI-Generated Image Naturalness (AGIN) database through a large-scale subjective study, collecting human opinions on the overall naturalness as well as perceptions from technical and rationality perspectives. The AGIN database verifies that naturalness is universally and disparately affected by both types of distortions. To automatically learn the naturalness of AGIs that aligns with human ratings, the authors propose the Joint Objective Image Naturalness evaluaTor (JOINT). JOINT imitates human reasoning in naturalness evaluation by jointly learning both technical and rationality perspectives. Experimental results demonstrate that JOINT outperforms baselines in providing more subjectively consistent results on naturalness assessment. In an appendix section, additional details are provided for the construction of the AGIN database. The authors explain that they collected multi-sourced images from five tasks to cover diverse appearances of naturalness issues. For text-to-image models, they chose 20 keywords and created 10 prompts for each keyword using GPT3.5. The generated prompts were manually examined for compliance before collecting AI-generated images. Furthermore, the authors propose JOINT++ as an extension to JOINT, incorporating additional features such as texture losses, blurs, artifacts, semantic information, attribute theme, layout, and aesthetic assessment elements into their model. Overall, this paper contributes to understanding and evaluating the naturalness of AGIs by presenting a benchmark database and objective evaluators that align with human opinions. The work has implications for future research in this field and encourages new research on assessing AGI naturalness via analysis of technical and rationality features; it also promotes development of better algorithms for assessing naturalness in AI-generated multimedia.