ResearchBenchmarksText-to-Image Generation on HPD Full rerun, n=3,200 v2Follow68.8PickScoreUG-FM55.69659.09862.565.902Jun 22, 2026Evaluation ResultsMethodMethodLinksPickScoreHPSCLIP ScoreAesthetic ScoreWilcoxon p (PickScore)UG-FMBackbone=SD3.5, Varian...Backbone=SD3.5, Variant=data-side, ηz=0.12026.0668.853.350.350.610PG-MAPBackbone=SD1.5, λ=0.1,...Backbone=SD1.5, λ=0.1, Scorer=PickScore2026.0658.855.852.355.27.9PG-MAPBackbone=SDXL, λ=0.05,...Backbone=SDXL, λ=0.05, Scorer=PickScore2026.0656.248.150.857.27.4