Image Generation on ImageNet-1k 1.0 (val)
4.1FIDVQ-KDCLIP
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| VQ-KDCLIPArchitecture=AR, #params=1.4B2024.11 | 4.1 | — | — | — | |
| FSQArchitecture=NAR, #params=227M2024.11 | 4.53 | — | — | — | |
| ViT-VQGANArchitecture=AR, #params=1.7B2024.11 | 5.3 | — | — | — | |
| MaskGITArchitecture=NAR, #params=227M2024.11 | 6.18 | — | — | — | |
| CVQ-VAEArchitecture=Diffusion, #params=400M2024.11 | 6.87 | — | — | — | |
| LDM-8-GArchitecture=Diffusion, #params=506M2024.11 | 7.76 | — | — | — | |
| RQ-VAEArchitecture=AR, #params=1.4B2024.11 | 8.71 | — | — | — | |
| MoVQArchitecture=NAR, #params=307M2024.11 | 8.78 | — | — | — | |
| Cosmos + Structured state-space regularizationAutoencoder type=Cosmos + Ours, Generative model=DiT-B/22026.04 | 11.2 | 128.91 | 72 | 45 | |
| Cosmos + EQVAEAutoencoder type=Cosmos + EQVAE, Generative model=DiT-B/22026.04 | 11.66 | 123.54 | 71 | 46 | |
| Cosmos + FT-SEAutoencoder type=Cosmos + FT-SE, Generative model=DiT-B/22026.04 | 12.01 | 117.97 | 71 | 46 | |
| Flux + Structured state-space regularizationAutoencoder type=Flux + Ours, Generative model=DiT-B/22026.04 | 13.33 | 106.61 | 71 | 47 | |
| CosmosAutoencoder type=Cosmos, Generative model=DiT-B/22026.04 | 13.52 | 112.48 | 68 | 47 | |
| Flux + FT-SEAutoencoder type=Flux + FT-SE, Generative model=DiT-B/22026.04 | 13.87 | 99.79 | 70 | 49 | |
| Flux + EQVAEAutoencoder type=Flux + EQVAE, Generative model=DiT-B/22026.04 | 13.95 | 97.81 | 69 | 49 | |
| VQGANArchitecture=AR, #params=1.4B2024.11 | 15.78 | — | — | — | |
| FluxAutoencoder type=Flux, Generative model=DiT-B/22026.04 | 16.08 | 89.32 | 68 | 49 |