Saturday 15 March 2025
The quest for fairness in AI-powered image generation has taken a significant step forward, as researchers have developed a new detector that can accurately identify gender bias in text-to-image models.
Text-to-image models have gained popularity in recent years due to their ability to generate high-quality images from natural language descriptions. However, these models have been shown to perpetuate harmful stereotypes and biases, particularly when it comes to gender. For instance, a study found that text-to-image models tend to associate male-dominated professions with masculine images, while female-dominated professions are often represented by feminine images.
To combat this issue, researchers have proposed various detectors aimed at uncovering gender bias in these models. However, existing detectors have been shown to be inaccurate or ineffective in detecting biases. This is because they rely on simplistic metrics, such as counting the number of male and female images generated, which can be misleading.
The new detector developed by researchers takes a more nuanced approach. It uses a combination of manual labeling and machine learning algorithms to identify gender bias in text-to-image models. The researchers created a dataset consisting of 6,000 images generated from three cutting-edge text-to-image models. During the human-labeling process, they found that all three models generate a portion (12.48% on average) of low-quality images, where human annotators cannot determine the gender of the person.
The study reveals that all three models show a preference for generating male images, with one model being significantly more biased than the others. Additionally, images generated using prompts containing professional descriptions (such as lawyer or doctor) show the most bias.
To test the effectiveness of their detector, the researchers evaluated seven existing gender bias detectors and found that none fully capture the actual level of bias in text-to-image models. Some detectors even overestimated the bias by up to 26.95%.
The new detector, called CLIP-Enhance, addresses these limitations by using a more sophisticated approach that considers multiple factors, including image quality, object recognition, and facial detection. The researchers found that CLIP-Enhance most accurately measures the gender bias in text-to-image models, with an average deviation of only 0.47% to 1.23%. Moreover, it filters out 82.91% of low-quality images.
The implications of this research are significant. As AI-powered image generation becomes increasingly prevalent in various industries, including healthcare, finance, and education, the need for fair and unbiased models grows more pressing.
Cite this article: “Detecting Gender Bias in Text-to-Image Models with CLIP-Enhance”, The Science Archive, 2025.
Ai-Powered Image Generation, Gender Bias, Text-To-Image Models, Fairness, Accuracy, Machine Learning Algorithms, Human Labeling, Object Recognition, Facial Detection, Clip-Enhance Detector
