The paper "New Job, New Gender? Measuring the Social Bias in Image Generation Models" by Wenxuan Wang et al. introduces BiasPainter, a novel evaluation framework designed to measure social bias in image generation models. These models, such as DALL-E and Midjourney, are capable of generating images from text but are often trained on large internet datasets, leading to the generation of content that perpetuates social stereotypes and biases. Previous research in this area has suffered from limitations such as low accuracy, reliance on extensive human labor, and lack of comprehensive analysis. BiasPainter addresses these issues by using a diverse range of seed images and prompts that are gender, race, and age-neutral. The framework compares the edited images to the original seed images, focusing on significant changes related to gender, race, and age. It evaluates six widely used image generation models, including Stable Diffusion, Midjourney, and DALL-E 2. The results show that BiasPainter can successfully trigger social bias in these models, achieving 90.8% accuracy in automatic bias detection, which is significantly higher than previous methods. The paper also includes a detailed methodology for collecting seed images, creating neutral prompts, generating images, assessing properties, and detecting bias. It provides insights into the nature and extent of biases within the models and offers a tool for evaluating bias mitigation strategies. The authors conclude that BiasPainter is effective in measuring social bias and can help mitigate biases in image generation models.The paper "New Job, New Gender? Measuring the Social Bias in Image Generation Models" by Wenxuan Wang et al. introduces BiasPainter, a novel evaluation framework designed to measure social bias in image generation models. These models, such as DALL-E and Midjourney, are capable of generating images from text but are often trained on large internet datasets, leading to the generation of content that perpetuates social stereotypes and biases. Previous research in this area has suffered from limitations such as low accuracy, reliance on extensive human labor, and lack of comprehensive analysis. BiasPainter addresses these issues by using a diverse range of seed images and prompts that are gender, race, and age-neutral. The framework compares the edited images to the original seed images, focusing on significant changes related to gender, race, and age. It evaluates six widely used image generation models, including Stable Diffusion, Midjourney, and DALL-E 2. The results show that BiasPainter can successfully trigger social bias in these models, achieving 90.8% accuracy in automatic bias detection, which is significantly higher than previous methods. The paper also includes a detailed methodology for collecting seed images, creating neutral prompts, generating images, assessing properties, and detecting bias. It provides insights into the nature and extent of biases within the models and offers a tool for evaluating bias mitigation strategies. The authors conclude that BiasPainter is effective in measuring social bias and can help mitigate biases in image generation models.