ResearchBenchmarksText-to-Image Generation on GenEval (Object Counting Compositionality)Text-to-Image Generation on GenEval (Object Counting Compositionality)Follow92Accuracy (2 objects)OSI80.5683.5386.589.47May 14, 2026Evaluation resultsMethodMethodLinksAccuracy (2 objects)Accuracy (3 objects)Accuracy (4 objects)Accuracy (5 objects)Accuracy (6 objects)Average Compositional AccuracyOSIBackbone=FLUXBackbone=FLUX2026.05927164404061TACABackbone=FLUXBackbone=FLUX2026.05896856312053OSIBackbone=SD3.5Backbone=SD3.52026.05898068463564PLADISBackbone=FLUXBackbone=FLUX2026.05877156312053TACABackbone=SD3.5Backbone=SD3.52026.05877455472658BaseBackbone=SD3.5Backbone=SD3.52026.05827259382455BaseBackbone=FLUXBackbone=FLUX2026.05816344291847