Search is not available for this dataset
# precision: int4 | gpu: NVIDIA GeForce RTX 3090 float64 0.71 10.8 | __index_level_0__ int64 512 1.02k | __index_level_1__ int64 4 20 | __index_level_2__ int64 0 2 |
|---|---|---|---|
10.81 | 1,024 | 20 | 0 |
10.13 | 1,024 | 20 | 1 |
10.15 | 1,024 | 20 | 2 |
4.32 | 1,024 | 8 | 0 |
4.32 | 1,024 | 8 | 1 |
4.32 | 1,024 | 8 | 2 |
2.37 | 1,024 | 4 | 0 |
2.37 | 1,024 | 4 | 1 |
2.37 | 1,024 | 4 | 2 |
2.74 | 512 | 20 | 0 |
2.73 | 512 | 20 | 1 |
2.73 | 512 | 20 | 2 |
1.22 | 512 | 8 | 0 |
1.22 | 512 | 8 | 1 |
1.22 | 512 | 8 | 2 |
0.71 | 512 | 4 | 0 |
0.71 | 512 | 4 | 1 |
0.71 | 512 | 4 | 2 |
dit-score results
Raw per-pair fidelity scores for quantized diffusion transformers, from the dit-score harness.
Study 1 (2026-07-16). Krea 2 Turbo, five quant formats vs BF16 reference, RTX 4090, 1024px, seed-locked, n=96 pairs per format. LPIPS (alex) + PSNR + ImageReward delta. Method, sample grids and analysis in the GitHub README.
- Downloads last month
- 30