Class-conditional image synthesis on ImageNet 256x256 1.0 (val)
6.18FIDMaskGIT
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| MaskGIT# params=227M, # steps=82022.02 | 6.18 | 182.1 | 80 | 51 | 63.14 | 84.45 | |
| BigGAN-deep# params=160M, # steps=12022.02 | 6.95 | 198.2 | 87 | 28 | 43.99 | 67.89 | |
| ADM# params=554M, # steps=2502022.02 | 10.94 | 101 | 69 | 63 | — | — | |
| Improved DDPM# params=280M, # steps=2502022.02 | 12.26 | — | 70 | 62 | — | — | |
| VQGAN# params=1.4B, # steps=2562022.02 | 15.78 | 78.3 | — | — | — | — | |
| VQGAN*# params=227M, # steps=256, setup=same architecture and setup as MaskGIT2022.02 | 18.65 | 80.4 | 78 | 26 | 53.1 | 76.18 | |
| VQVAE-2# params=13.5B, # steps=51202022.02 | 31.11 | 45 | 36 | 57 | 54.83 | 77.59 | |
| DCTransformer# params=738M, # steps=>10242022.02 | 36.51 | — | 36 | 67 | — | — |