ResearchBenchmarksText-to-Image Generation on Compositional multi-subject Prompt FamilyText-to-Image Generation on Compositional multi-subject Prompt FamilyFollow0.045Gender-debiasing Parity GapEquiSteer0.0350.10250.170.2375Jul 1, 2026Evaluation resultsMethodMethodLinksGender-debiasing Parity GapEquiSteerBackbone=SDXLBackbone=SDXL2026.070.045EquiSteerBackbone=SANABackbone=SANA2026.070.155VanillaBackbone=SDXLBackbone=SDXL2026.070.262VanillaBackbone=SANABackbone=SANA2026.070.295