Text-to-Image Generation on GenEval Text Rendering and Human Preference Evaluation Suite (test)
0.95GenEval ScoreSD3.5-M + FlowGRPO
Evaluation results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| SD3.5-M + FlowGRPOBase Model=SD3.5-M, Post-training=FlowGRPO, Optimization Reward=GenEval2026.03 | 0.95 | 65 | 29.3 | 5.32 | 22.51 | 27.2 | 1.06 | 3.18 | |
| SD3.5-M + FlowGRPO + ARCBase Model=SD3.5-M, Post-training=FlowGRPO + ARC, Optimization Reward=GenEval2026.03 | 0.92 | 71 | 29.4 | 5.35 | 22.5 | 27.7 | 1.06 | 3.26 | |
| SD3.5-M + FlowGRPO + ARCBase Model=SD3.5-M, Post-training=FlowGRPO + ARC, Optimization Reward=PickScore2026.03 | 0.77 | 70 | 28.7 | 5.63 | 22.73 | 28.6 | 1.07 | 3.26 | |
| SD3.5-M + FlowGRPO + ARCBase Model=SD3.5-M, Post-training=FlowGRPO + ARC, Optimization Reward=OCR2026.03 | 0.72 | 89 | 29.1 | 5.39 | 22.45 | 28.4 | 0.97 | 3.19 | |
| SD3.5-LBase Model=SD3.5-L2026.03 | 0.71 | 68 | 28.9 | 5.5 | 22.91 | 28.8 | 0.96 | 3.25 | |
| SD3.5-M + ARCBase Model=SD3.5-M, Post-training=ARC2026.03 | 0.71 | 67 | 28.8 | 5.39 | 22.41 | 27.8 | 0.87 | 3.11 | |
| SD3.5-M + FlowGRPOBase Model=SD3.5-M, Post-training=FlowGRPO, Optimization Reward=OCR2026.03 | 0.67 | 92 | 29 | 5.32 | 22.41 | 28 | 0.95 | 3.14 | |
| SD3.5-MBase Model=SD3.5-M2026.03 | 0.65 | 61 | 28.2 | 5.36 | 22.34 | 27.9 | 0.84 | 3.08 | |
| SD-XLBase Model=SD-XL2026.03 | 0.55 | 14 | 28.7 | 5.6 | 22.42 | 28 | 0.76 | 2.93 | |
| SD3.5-M + FlowGRPOBase Model=SD3.5-M, Post-training=FlowGRPO, Optimization Reward=PickScore2026.03 | 0.54 | 68 | 27.8 | 5.9 | 23.5 | 31.4 | 1.26 | 3.37 |