Image Reconstruction on ImageNet (test)
0.686FIDSber-MoVQGAN 270M
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| Sber-MoVQGAN 270MLatent size=32x32, Num Z=16384, Train steps=1,330,0002023.10 | 0.686 | 0.741 | 27.04 | 0.0393 | |
| Sber-MoVQGAN 102MLatent size=32x32, Num Z=16384, Train steps=2,360,0002023.10 | 0.776 | 0.737 | 26.89 | 0.0398 | |
| Sber-MoVQGAN 67MLatent size=32x32, Num Z=16384, Train steps=2,000,0002023.10 | 0.965 | 0.725 | 26.45 | 0.0415 | |
| KL CompVisLatent size=32x32, Train steps=246,8032023.10 | 0.968 | 0.692 | 25.11 | 0.0474 | |
| Mo-VQGAN*Latent size=16x16x4, Num Z=1024, Train steps=40 epochs2023.10 | 1.12 | 0.673 | 22.42 | — | |
| VIT-VQGAN*Latent size=32x32, Num Z=8192, Train steps=500,0002023.10 | 1.28 | — | — | — | |
| VQ CompVisLatent size=32x32, Num Z=16384, Train steps=971,0432023.10 | 1.34 | 0.65 | 23.85 | 0.0533 | |
| Sber-MoVQGAN 67MLatent size=32x32, Num Z=1024, Train steps=5,000,0002023.10 | 1.34 | 0.704 | 25.68 | 0.0451 | |
| Sber-VQGANLatent size=32x32, Num Z=8192, Train steps=1 epoch2023.10 | 1.44 | 0.682 | 24.31 | 0.0503 | |
| RQ-VAE*Latent size=8x8x16, Num Z=16384, Train steps=10 epochs2023.10 | 1.83 | — | — | — |