Text-to-image models have made rapid progress in photorealism, semantic fidelity, and aesthetics. Yet a new benchmark, HealBench, argues that a crucial human-centered quality has been left out: whether the images feel psychologically comforting. HealBench is a fine-grained benchmark designed to evaluate this dimension in both generation and editing tasks.
The paper notes that despite advances in technical metrics, psychological comfort remains understudied. HealBench aims to fill that gap by providing a structured way to assess how comforting generated images are to viewers. The benchmark is presented as a complement to existing evaluation frameworks, which typically focus on objective qualities rather than subjective emotional responses.
The work highlights a broader shift toward human-centered evaluation in AI, where the emotional impact of generated content is considered alongside technical performance. While the abstract does not detail the benchmark's construction or results, it signals a growing interest in making text-to-image systems more attuned to psychological well-being.