# TWIGMA: A dataset of AI-Generated Images with Metadata From Twitter

**Authors:** Yiqun Chen, James Zou  
**arXiv:** [2306.08310](https://arxiv.org/abs/2306.08310)  
**Date:** 2023-06-14  
**Summary page:** https://www.summarizepaper.com/en/paper/2306.08310v1/

## Key points

- Recent progress in generative artificial intelligence (gen-AI) has revolutionized the creation of photo-realistic and artistically-inspiring images
- TWIGMA is an extensive dataset comprising over 800,000 gen-AI images collected from January 2021 to March 2023 on Twitter, along with associated metadata
- Gen-AI images possess distinctive characteristics and exhibit lower variability compared to natural images and human artwork
- There is an inverse correlation between the similarity of a gen-AI image to natural images and the number of likes it receives
- The similarity measure can be utilized to identify human images that served as inspiration for gen-AI creations
- Over time, users have increasingly shared artistically sophisticated content such as intricate human portraits while showing decreased interest in simple subjects like natural scenes and animals
- Figure 1 showcases the curation process of TWIGMA, the growth in tweets featuring generative AI images over time, and presents a wordcloud highlighting prevalent keywords extracted from the tweets
- Section 2 reviews relevant work on image datasets and evaluation metrics for model novelty and variation
- Section 3 outlines the data collection and analysis process
- Empirical results addressing the research questions are presented in Section 4.

## Summary

Recent progress in generative artificial intelligence (gen-AI) has revolutionized the creation of photo-realistic and artistically-inspiring images, making them easily accessible to millions of users online. To gain insights into how people utilize gen-AI models like DALLE and StableDiffusion, it is crucial to understand the themes, contents, and variations present in these AI-generated photos. In this paper, we introduce TWIGMA (TWItter Generative-ai images with MetadatA), an extensive dataset comprising over 800,000 gen-AI images collected from January 2021 to March 2023 on Twitter, along with associated metadata such as tweet text, creation date, and number of likes. To analyze TWIGMA comprehensively, we compare it with natural images and human artwork. Our findings reveal that gen-AI images possess distinctive characteristics and exhibit lower variability on average compared to their non-gen-AI counterparts. Interestingly, we observe an inverse correlation between the similarity of a gen-AI image to natural images and the number of likes it receives. Moreover, we demonstrate that the similarity measure can be utilized to identify human images that served as inspiration for gen-AI creations. Furthermore, our analysis reveals a longitudinal shift in the themes of AI-generated images shared on Twitter. Over time, users have increasingly shared artistically sophisticated content such as intricate human portraits while showing decreased interest in simple subjects like natural scenes and animals. The paper also provides additional context through Figure 1. It showcases the curation process of TWIGMA resulting in approximately 800,000 images posted from January 2021 to March 2023. The figure also depicts the steady growth in tweets featuring generative AI images over time and presents a wordcloud highlighting prevalent keywords extracted from the tweets. In terms of organization, the rest of the paper is structured as follows: Section 2 reviews relevant work on image datasets and evaluation metrics for model novelty and variation. Section 3 outlines the data collection and analysis process. Empirical results addressing the research questions are presented in Section 4.