Human preference has become central to how text-to-image models are evaluated, guided, fine-tuned, and ranked. A new arXiv paper (2610.28195v1) argues that the susceptibility of these preferences to bias in image generation remains underexplored, and identifies a key issue: attractive faces.
According to the abstract, this visual bias can skew text-to-image reward models and leaderboards. Because reward models are trained on human preferences, systematic preferences for attractive faces may distort which outputs are judged as higher quality, potentially affecting both automated evaluation and model guidance.
Since only one source is available, there is no independent comparison to report. The paper's contribution is to flag this bias as an important but under-examined factor in the growing use of human preference in text-to-image systems.