Preference ≠ Tattooability: Auditing Reward Models for AI Tattoo Generation

Abstract

AI tattoo generation is usually ranked with the same reward models that rank general text-to-image outputs. Those models predict which image a crowd annotator prefers. A tattoo that will be worn also has to survive a needle, a gap that heals shut, and ink that spreads. We audited that gap with deterministic, physically scaled probes and with CLIPScore, PickScore, HPSv2.1 and ImageReward, on tattoo-subject pairs from HPDv2 and ImageRewardDB, on controlled defects, and on a print-scale ink-spread stress test. On HPDv2 tattoo-subject pairs, a-priori probes agree with human preference below chance, and that contrast remains after a Holm correction across scorers. Nothing here ranks a commercial AI tattoo generator. Competing interest: the author operates tatany.app; that product was not evaluated or ranked.