Text-to-Image Generation on GenEval (val)
90GenEval ScoreAPEX
Evaluation Results
| Method | Links | |
|---|---|---|
| APEXNFEs=2, Throughput (samples/s)=3.30, Latency (s)=0.45, Params (B)=20, Training Protocol=Full tuning2026.04 | 90 | |
| APEXNFEs=2, Throughput (samples/s)=3.17, Latency (s)=0.47, Backbone Params (B)=20, Training Protocol=LoRA, LoRA rank=64, Trainable Params (B)=0.42026.04 | 89 | |
| APEXNFEs=1, Throughput (samples/s)=3.27, Latency (s)=0.39, Backbone Params (B)=20, Training Protocol=LoRA, LoRA rank=64, Trainable Params (B)=0.42026.04 | 89 | |
| APEXNFEs=1, Throughput (samples/s)=3.50, Latency (s)=0.34, Params (B)=20, Training Protocol=Full tuning2026.04 | 89 | |
| APEXNFEs=1, Throughput (samples/s)=3.29, Latency (s)=0.39, Backbone Params (B)=20, Training Protocol=LoRA, LoRA rank=32, Trainable Params (B)=0.22026.04 | 88 | |
| TwinFlowNFEs=2, Throughput (samples/s)=3.15, Latency (s)=0.48, Backbone Params (B)=20, Training Protocol=LoRA, LoRA rank=64, Trainable Params (B)=0.42026.04 | 87 | |
| APEXNFEs=2, Throughput (samples/s)=3.21, Latency (s)=0.49, Backbone Params (B)=20, Training Protocol=LoRA, LoRA rank=32, Trainable Params (B)=0.22026.04 | 87 | |
| TwinFlowNFEs=1, Throughput (samples/s)=3.29, Latency (s)=0.40, Backbone Params (B)=20, Training Protocol=LoRA, LoRA rank=64, Trainable Params (B)=0.42026.04 | 86 | |
| Qwen-Image-LightningNFEs=2, Throughput (samples/s)=3.15, Latency (s)=0.48, Backbone Params (B)=20, Training Protocol=LoRA, LoRA rank=64, Trainable Params (B)=0.42026.04 | 85 | |
| APEXNFEs=2, Throughput (samples/s)=5.72, Latency (s)=0.23, Params (B)=1.6, Training Protocol=Full tuning2026.04 | 85 | |
| Qwen-Image-LightningNFEs=1, Throughput (samples/s)=3.29, Latency (s)=0.40, Backbone Params (B)=20, Training Protocol=LoRA, LoRA rank=64, Trainable Params (B)=0.42026.04 | 85 | |
| APEXNFEs=2, Throughput (samples/s)=6.50, Latency (s)=0.25, Params (B)=0.6, Training Protocol=Full tuning2026.04 | 84 | |
| APEXNFEs=1, Throughput (samples/s)=7.30, Latency (s)=0.20, Params (B)=0.6, Training Protocol=Full tuning2026.04 | 84 | |
| APEXNFEs=1, Throughput (samples/s)=6.84, Latency (s)=0.20, Params (B)=1.6, Training Protocol=Full tuning2026.04 | 84 | |
| RCGMNFEs=2, Throughput (samples/s)=3.15, Latency (s)=0.48, Backbone Params (B)=20, Training Protocol=LoRA, LoRA rank=64, Trainable Params (B)=0.42026.04 | 82 | |
| Sana-SprintNFEs=2, Throughput (samples/s)=5.68, Latency (s)=0.24, Params (B)=1.6, Training Protocol=Full tuning2026.04 | 77 | |
| Sana-SprintNFEs=2, Throughput (samples/s)=6.46, Latency (s)=0.25, Params (B)=0.6, Training Protocol=Full tuning2026.04 | 76 | |
| Sana-SprintNFEs=1, Throughput (samples/s)=6.71, Latency (s)=0.21, Params (B)=1.6, Training Protocol=Full tuning2026.04 | 76 | |
| Sana-SprintNFEs=1, Throughput (samples/s)=7.22, Latency (s)=0.21, Params (B)=0.6, Training Protocol=Full tuning2026.04 | 72 | |
| FLUX-schnellNFEs=2, Throughput (samples/s)=0.92, Latency (s)=1.15, Params (B)=12.0, Training Protocol=Full tuning2026.04 | 71 | |
| FLUX-schnellNFEs=1, Throughput (samples/s)=1.58, Latency (s)=0.68, Params (B)=12.0, Training Protocol=Full tuning2026.04 | 69 | |
| SDXL-DMD2NFEs=1, Throughput (samples/s)=3.36, Latency (s)=0.32, Params (B)=0.9, Training Protocol=Full tuning, Model Strategy=Distinct models per NFE2026.04 | 59 | |
| SDXL-DMD2NFEs=2, Throughput (samples/s)=2.89, Latency (s)=0.40, Params (B)=0.9, Training Protocol=Full tuning, Model Strategy=Distinct models per NFE2026.04 | 58 | |
| PCMNFEs=2, Throughput (samples/s)=2.62, Latency (s)=0.56, Params (B)=0.9, Training Protocol=Full tuning, Model Strategy=Distinct models per NFE2026.04 | 55 | |
| SD3.5-TurboNFEs=2, Throughput (samples/s)=1.61, Latency (s)=0.68, Params (B)=8.0, Training Protocol=Full tuning2026.04 | 53 | |
| RCGMNFEs=1, Throughput (samples/s)=3.29, Latency (s)=0.40, Backbone Params (B)=20, Training Protocol=LoRA, LoRA rank=64, Trainable Params (B)=0.42026.04 | 52 | |
| SD3.5-TurboNFEs=1, Throughput (samples/s)=2.48, Latency (s)=0.45, Params (B)=8.0, Training Protocol=Full tuning2026.04 | 51 | |
| PixArt-DMDNFEs=1, Throughput (samples/s)=4.26, Latency (s)=0.25, Params (B)=0.6, Training Protocol=Full tuning, Model Strategy=Distinct models per NFE2026.04 | 45 | |
| SDXL-LCMNFEs=2, Throughput (samples/s)=2.89, Latency (s)=0.40, Params (B)=0.9, Training Protocol=Full tuning2026.04 | 44 | |
| PixArt-LCMNFEs=2, Throughput (samples/s)=3.52, Latency (s)=0.31, Params (B)=0.6, Training Protocol=Full tuning2026.04 | 42 | |
| PCMNFEs=1, Throughput (samples/s)=3.16, Latency (s)=0.40, Params (B)=0.9, Training Protocol=Full tuning, Model Strategy=Distinct models per NFE2026.04 | 42 | |
| PixArt-LCMNFEs=1, Throughput (samples/s)=4.26, Latency (s)=0.25, Params (B)=0.6, Training Protocol=Full tuning2026.04 | 41 | |
| SDXL-LCMNFEs=1, Throughput (samples/s)=3.36, Latency (s)=0.32, Params (B)=0.9, Training Protocol=Full tuning2026.04 | 28 |