ResearchBenchmarksText-to-Image Generation on HPD 4-specialization sweep, n=800 v2Follow69.5PickScoreUG-FM51.50856.17960.8565.521Jun 22, 2026Evaluation ResultsMethodMethodLinksPickScoreHPSCLIP ScoreAesthetic ScoreWilcoxon p-value (PS)UG-FMBackbone=SD3.5Backbone=SD3.52026.0669.554.954.648.65.5MAP-czBackbone=SDXLBackbone=SDXL2026.0657.649.851.957.11.4Reward-zBackbone=SD1.5Backbone=SD1.52026.0657.15752.856.11.2MAP-czBackbone=SD1.5Backbone=SD1.52026.0656.655.851.755.62MAP-cBackbone=SD1.5Backbone=SD1.52026.0652.249.449.644.20.347