Text-to-Image Generation on DCI
22.1FIDBaseline
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| BaselineBackbone=FLUX-DiT, W/A=BF162025.07 | 22.1 | 0.618 | 0.451 | 18.03 | — | — | — | |
| BaselineBackbone=SD3.5-DiT, W/A=FP162025.07 | 22.17 | 0.685 | 0.451 | 18.05 | — | — | — | |
| SegQuant-GBackbone=FLUX-DiT, W/A=8/8(int)2025.07 | 22.19 | 0.59 | 0.451 | 18.02 | 0.131 | 18.97 | 0.737 | |
| SegQuant-ABackbone=FLUX-DiT, W/A=8/8(int)2025.07 | 22.28 | 0.592 | 0.452 | 18.01 | 0.149 | 18.29 | 0.715 | |
| SVDQuantBackbone=FLUX-DiT, W/A=4/8(int)2025.07 | 22.57 | 0.575 | 0.442 | 18.03 | 0.224 | 16.19 | 0.635 | |
| SegQuant-GBackbone=FLUX-DiT, W/A=4/8(int)2025.07 | 23.31 | 0.566 | 0.446 | 18.03 | 0.219 | 16.31 | 0.639 | |
| SegQuant-GBackbone=SD3.5-DiT, W/A=4/8(int)2025.07 | 23.5 | 0.576 | 0.435 | 18.14 | 0.438 | 12.25 | 0.497 | |
| Q-DiffusionBackbone=FLUX-DiT, W/A=8/8(int)2025.07 | 23.67 | 0.513 | 0.449 | 18.01 | 0.298 | 14.4 | 0.563 | |
| SegQuant-GBackbone=SDXL-UNet, W/A=8/8(int)2025.07 | 24.12 | 0.504 | 0.416 | 17.94 | 0.123 | 22.11 | 0.763 | |
| SegQuant-ABackbone=SD3.5-DiT, W/A=8/8(int)2025.07 | 24.13 | 0.639 | 0.44 | 18.07 | 0.407 | 12.74 | 0.521 | |
| BaselineBackbone=SDXL-UNet, W/A=FP162025.07 | 24.27 | 0.495 | 0.417 | 17.94 | — | — | — | |
| PTQ4DiTBackbone=SDXL-UNet, W/A=8/8(int)2025.07 | 24.34 | 0.472 | 0.415 | 17.94 | 0.197 | 19.63 | 0.686 | |
| SegQuant-ABackbone=SDXL-UNet, W/A=8/8(int)2025.07 | 24.44 | 0.487 | 0.416 | 17.94 | 0.132 | 21.69 | 0.752 | |
| SegQuant-GBackbone=SD3.5-DiT, W/A=8/8(int)2025.07 | 24.58 | 0.639 | 0.447 | 18.06 | 0.412 | 12.4 | 0.516 | |
| SVDQuantBackbone=SD3.5-DiT, W/A=4/8(int)2025.07 | 25.22 | 0.518 | 0.43 | 18.14 | 0.446 | 12.34 | 0.495 | |
| SegQuant-GBackbone=SDXL-UNet, W/A=8/8(fp)2025.07 | 26.1 | 0.474 | 0.405 | 17.89 | 0.092 | 23.82 | 0.793 | |
| SegQuant-ABackbone=SDXL-UNet, W/A=8/8(fp)2025.07 | 26.13 | 0.475 | 0.406 | 17.89 | 0.101 | 23.34 | 0.782 | |
| PTQ4DiTBackbone=FLUX-DiT, W/A=8/8(int)2025.07 | 26.39 | 0.438 | 0.472 | 17.97 | 0.552 | 9.22 | 0.321 | |
| Q-DiffusionBackbone=SDXL-UNet, W/A=8/8(fp)2025.07 | 26.98 | 0.474 | 0.406 | 17.89 | 0.102 | 23.28 | 0.78 | |
| PTQ4DiTBackbone=SD3.5-DiT, W/A=8/8(int)2025.07 | 27.01 | 0.485 | 0.43 | 18.08 | 0.445 | 12.33 | 0.492 | |
| TACBackbone=SD3.5-DiT, W/A=8/8(int)2025.07 | 28.33 | 0.43 | 0.432 | 18.1 | 0.461 | 12.11 | 0.478 | |
| PTQ4DiTBackbone=SD3.5-DiT, W/A=4/8(int)2025.07 | 61.68 | -0.314 | 0.403 | 18.21 | 0.566 | 10.94 | 0.386 | |
| PTQDBackbone=SD3.5-DiT, W/A=8/8(int)2025.07 | 65.42 | -0.454 | 0.416 | 18.08 | 0.582 | 10 | 0.296 | |
| PTQ4DiTBackbone=FLUX-DiT, W/A=4/8(int)2025.07 | 72.86 | -0.393 | 0.418 | 17.88 | 0.592 | 10.98 | 0.465 | |
| Q-DiffusionBackbone=SDXL-UNet, W/A=8/8(int)2025.07 | 73.39 | -1.466 | 0.402 | 17.65 | 0.552 | 13.64 | 0.448 | |
| Q-DiffusionBackbone=SD3.5-DiT, W/A=8/8(int)2025.07 | 143.36 | -1.973 | 0.46 | 17.92 | 0.691 | 7.8 | 0.254 |