Text-to-Image Generation on MS-COCO 5K 2017 (val)
23.21FIDClockwork (Efficient UNet backbone)
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Clockwork (Efficient UNet backbone)Resolution=512x512, Sampling Steps=8, Scheduler=DPM++, TFLOPs=5.9, Latency (GPU)=213 ms, Latency (Phone)=1196 ms2023.12 | 23.21 | 0.296 | — | — | — | — | — | |
| Clockwork (SD UNet backbone)Resolution=512x512, Sampling Steps=8, Scheduler=DPM++, TFLOPs=7.3, Latency (GPU)=341 ms, Latency (Phone)=3176 ms2023.12 | 24.11 | 0.295 | — | — | — | — | — | |
| SnapFusionResolution=512x512, Sampling Steps=8, Scheduler=DPM++, TFLOPs=4.0, Latency (GPU)=185 ms2023.12 | 24.2 | 0.3 | — | — | — | — | — | |
| Efficient UNetResolution=512x512, Sampling Steps=8, Scheduler=DPM++, TFLOPs=9.5, Latency (GPU)=330 ms, Latency (Phone)=1960 ms2023.12 | 24.22 | 0.302 | — | — | — | — | — | |
| Clockwork (Distilled Efficient UNet backbone)Resolution=512x512, Sampling Steps=8, Scheduler=DPM++, TFLOPs=2.9, Latency (GPU)=154 ms, Latency (Phone)=598 ms2023.12 | 24.45 | 0.295 | — | — | — | — | — | |
| SD UNetResolution=512x512, Sampling Steps=8, Scheduler=DPM++, TFLOPs=10.8, Latency (GPU)=454 ms, Latency (Phone)=3968 ms2023.12 | 24.64 | 0.3 | — | — | — | — | — | |
| Distilled Efficient UNetResolution=512x512, Sampling Steps=8, Scheduler=DPM++, TFLOPs=4.7, Latency (GPU)=240 ms, Latency (Phone)=980 ms2023.12 | 25.75 | 0.297 | — | — | — | — | — | |
| Meng et al.Resolution=512x512, Sampling Steps=8, Scheduler=DPM++, TFLOPs=6.4, Latency (GPU)=320 ms2023.12 | 26.9 | 0.3 | — | — | — | — | — | |
| CCMBase Model=SDXL2024.12 | 28.9 | 0.326 | — | — | — | — | — | |
| BK-SDM-BaseResolution=512x512, Sampling Steps=8, Scheduler=DPM++, TFLOPs=8.4, Latency (GPU)=348 ms2023.12 | 29.26 | 0.291 | — | — | — | — | — | |
| InstaFlowResolution=512x512, Sampling Steps=1, TFLOPs=0.8, Latency (GPU)=40 ms2023.12 | 29.3 | 0.283 | — | — | — | — | — | |
| BK-SDM-SmallResolution=512x512, Sampling Steps=8, Scheduler=DPM++, TFLOPs=8.2, Latency (GPU)=336 ms2023.12 | 29.48 | 0.272 | — | — | — | — | — | |
| PCMBase Model=SDXL2024.12 | 29.89 | 0.3247 | — | — | — | — | — | |
| Hyper-SDBase Model=SDXL2024.12 | 30.38 | 0.321 | — | — | — | — | — | |
| Tar-1.5B + LANTERNResolution=1024, Model Size=1.5B2026.01 | 31.1 | — | 1.42 | 28.5 | — | — | — | |
| BK-SDM-TinyResolution=512x512, Sampling Steps=8, Scheduler=DPM++, TFLOPs=7.8, Latency (GPU)=313 ms2023.12 | 31.48 | 0.268 | — | — | — | — | — | |
| Tar-1.5BResolution=1024, Model Size=1.5B2026.01 | 32.4 | — | 1 | 29.5 | — | — | — | |
| Tar-1.5B + ZipAR-16Resolution=1024, Model Size=1.5B2026.01 | 32.4 | — | 3.65 | 29.6 | — | — | — | |
| Tar-1.5B + EAGLE-2Resolution=1024, Model Size=1.5B2026.01 | 32.4 | — | 0.78 | 29.5 | — | — | — | |
| CCMBase Model=SD32024.12 | 32.54 | 0.3242 | — | — | — | — | — | |
| Tar-1.5B + LANTERNResolution=512, Model Size=1.5B2026.01 | 32.7 | — | 1.08 | 27.7 | — | — | — | |
| Tar-1.5B + MuLo-SDResolution=1024, Model Size=1.5B, Parallelism=4x2026.01 | 32.8 | — | 1.68 | 28.4 | — | — | — | |
| Tar-1.5BResolution=512, Model Size=1.5B2026.01 | 33 | — | 1 | 28.6 | — | — | — | |
| Tar-1.5B + ZipAR-16Resolution=512, Model Size=1.5B2026.01 | 33 | — | 1.88 | 28.5 | — | — | — | |
| Tar-1.5B + EAGLE-2Resolution=512, Model Size=1.5B2026.01 | 33 | — | 0.72 | 28.6 | — | — | — | |
| PCMBase Model=SD32024.12 | 33.22 | 0.3234 | — | — | — | — | — | |
| Tar-1.5B + MuLo-SDResolution=512, Model Size=1.5B, Parallelism=2x2026.01 | 33.7 | — | 1.22 | 27.8 | — | — | — | |
| LCMBase Model=SD32024.12 | 35.62 | 0.3232 | — | — | — | — | — | |
| PCPOModel=FLUX2025.09 | 40.38 | — | — | — | 438.88 | 11.9 | 0.5708 | |
| BaselineModel=FLUX2025.09 | 46.23 | — | — | — | 539.83 | 12.66 | 0.5736 | |
| OriginalBase Model=SDXL2024.12 | 70.28 | 0.3041 | — | — | — | — | — | |
| PCPOModel=SD1.42025.09 | 84.74 | — | — | — | 1,035.45 | 7.5 | 0.4894 | |
| BaselineModel=SD1.42025.09 | 90.34 | — | — | — | 1,078.42 | 7.61 | 0.4948 | |
| OriginalBase Model=SD32024.12 | 99.61 | 0.2809 | — | — | — | — | — |