Unconditional Image Synthesis on CelebA-HQ 256 x 256
3.15FIDRDM
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| RDMTraining iterations=50M trained images, Number of generated samples=30,0002023.09 | 3.15 | 0.77 | 0.55 | |
| StyleSwinbCR=true2021.12 | 3.25 | — | — | |
| StyleSwinTraining iterations=820M trained images, Number of generated samples=30,0002023.09 | 3.25 | — | — | |
| HiT-BbCR=true2021.12 | 3.39 | — | — | |
| LDM-4Sampling steps=5002021.12 | 5.11 | 0.72 | 0.49 | |
| LDM-4Number of generated samples=30,0002023.09 | 5.11 | 0.72 | 0.49 | |
| WaveDiffNumber of generated samples=30,0002023.09 | 5.94 | — | 0.37 | |
| UDM2021.12 | 7.16 | — | — | |
| LSGM2021.12 | 7.22 | — | — | |
| LSGMNumber of generated samples=30,0002023.09 | 7.22 | — | — | |
| PGGAN2021.12 | 8 | — | — | |
| PG-GANbCR=false2021.12 | 8.03 | — | — | |
| TransGANbCR=true2021.12 | 9.6 | — | — | |
| VQGAN+T. (k=400)2021.12 | 10.2 | — | — | |
| VQGANbCR=false2021.12 | 10.7 | — | — | |
| DC-VAE2021.12 | 15.8 | — | — |