Compositional text-to-image generation on GenEval official (test) benchmark leaderboard · SOTA2 Research

Compositional text-to-image generation on GenEval official (test)

100Single Object AccuracyFlow-GRPO

Evaluation Results

Method	Links
Flow-GRPO2026.05		100	99	95	93	98	82	94
CGPO2026.05		100	99	96	94	99	89	96
Janus-Pro2026.05		99	89	59	90	79	66	80
GPT-4o2026.05		99	92	85	92	75	61	84
SANA-1.52026.05		99	93	86	84	59	65	81
CNS2026.05		98.8	79.4	71.4	80.3	16.4	42	64.7
ODE2026.05		98.8	78.4	69.9	82.6	15.2	41.3	64.3
SDE2026.05		98.4	78.9	68.5	76.8	15.4	43.3	63.5
SDXL2026.05		98	74	39	85	15	23	55
FLUX.12026.05		98	81	74	79	22	45	66
SD3.52026.05		98	89	73	83	34	47	71
SD3.52026.05		98	78	50	81	24	52	63
DALLE-32026.05		96	87	47	83	43	45	67