TASTE: A Designer-Annotated Multi-Dimensional Preference Dataset for AI-Generated Graphic Design
The article discusses the release of TASTE, a dataset designed to evaluate AI-generated graphic design. It includes ratings from professional designers on various criteria, highlighting the limitations of existing models in accurately assessing design quality. The findings suggest that while some models perform better than others, none fully meet the standards set by human designers.
- ▪TASTE is a dataset that includes ratings from ten professional designers on outputs from four text-to-image models across nine criteria.
- ▪The dataset aims to address the shortcomings of existing models that rely on single overall verdicts for graphic design evaluation.
- ▪A small pairwise-difference head trained on TASTE achieved a score of 0.611, closing the gap to the single-rater ceiling.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Computer Vision and Pattern Recognition arXiv:2605.20731 (cs) [Submitted on 20 May 2026] Title:TASTE: A Designer-Annotated Multi-Dimensional Preference Dataset for AI-Generated Graphic Design Authors:Haonan Zhu, Elad Hirsch, Alexandria Minetti, Allison Nulty, Purvanshi Mehta View a PDF of the paper titled TASTE: A Designer-Annotated Multi-Dimensional Preference Dataset for AI-Generated Graphic Design, by Haonan Zhu and 4 other authors View PDF HTML (experimental) Abstract:Text-to-image models produce graphic design at production scale, but their supervision comes from photo-style preference data with a single overall verdict per comparison.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
