A new research paper from arXiv highlights that human preference in evaluating text-to-image models is significantly skewed by the attractiveness of faces depicted in generated images. This bias can override judgments of image quality and even explicit user prompts, leading to leaderboards that favor synthetic attractive faces over real ones. The study found that attractiveness is a strong predictor of scores across multiple human preference models, suggesting a need for preference models that can better distinguish image quality from cognitive biases. AI
IMPACT Highlights the need for more robust AI evaluation metrics that account for human cognitive biases, potentially impacting future model development and deployment.
RANK_REASON Academic paper on AI model evaluation bias. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
