Summary of Survey Of Bias in Text-to-image Generation: Definition, Evaluation, and Mitigation, by Yixin Wan et al.
Survey of Bias In Text-to-Image Generation: Definition, Evaluation, and Mitigation
by Yixin Wan, Arjun Subramonian, Anaelia Ovalle, Zongyu Lin, Ashima Suvarna, Christina Chance, Hritik Bansal, Rebecca Pattichis, Kai-Wei Chang
First submitted to arxiv on: 1 Apr 2024
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: Artificial Intelligence (cs.AI); Computers and Society (cs.CY)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The recent advancement of large and powerful Text-to-Image (T2I) generation models, such as OpenAI’s DALLE-3 and Google’s Gemini, enables users to generate high-quality images from textual prompts. However, these models have been found to exhibit conspicuous social bias in generated images, leading to allocational and representational harms in society, further marginalizing minority groups. Recent works have investigated different dimensions of bias in T2I systems, but a systematic review is lacking, hindering an understanding of current progress and research gaps. The paper presents the first extensive survey on bias in T2I generative models, reviewing prior studies on gender, skintone, and geo-cultural biases. The authors found that while gender and skintone biases are widely studied, geo-cultural bias remains under-explored; most works on gender and skintone bias investigated occupational association, while other aspects are less frequently studied; almost all gender bias works overlook non-binary identities; evaluation datasets and metrics are scattered, with no unified framework for measuring biases; and current mitigation methods fail to resolve biases comprehensively. The authors point out future research directions that contribute to human-centric definitions, evaluations, and mitigation of biases. The paper highlights the importance of studying biases in T2I systems and encourages future efforts to holistically understand and tackle biases, building fair and trustworthy T2I technologies for everyone. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary T2I models can create high-quality images from text prompts. However, they often show social bias, making people from certain groups feel left out or misrepresented. Some studies have looked at this problem, but we don’t know much about what’s been done or where we need to go next. This paper takes a close look at the work that’s already been done on understanding and fixing these biases. We found that most research has focused on gender and skin tone biases, but there’s still a lot to learn about geo-cultural bias. The methods used to study and fix these biases are not always good or comprehensive. We also found that many studies ignore non-binary people or leave them out of the discussion. The paper encourages more research into fixing biases in T2I models so we can have fair and trustworthy technology for everyone. |
Keywords
» Artificial intelligence » Gemini
