ResearchBenchmarksCompositional Image Generation on EvalGenFollow0.96EvalGen ScoreOP-GRPO0.61680.70590.7950.8841Apr 5, 2026Evaluation ResultsMethodMethodLinksEvalGen ScoreAesthetic ScoreDeQA ScoreImage Reward ScorePickScoreUnified Reward ScoreOP-GRPOStep=BestStep=Best2026.040.965.173.981.1322.423.57Flow-GRPOStep=BestStep=Best2026.040.955.1841.1222.373.51OP-GRPOStep=1020Step=10202026.040.885.224.021.2122.353.49Flow-GRPOStep=1020Step=10202026.040.685.274.051.3722.243.35SD3.5-MStep=—Step=—2026.040.635.314.070.9222.183.29