Class-conditional Image Generation on ImageNet 256x256 2012 (train val)
1.26FID (w/o G)EDM2-L + DDO (downsampled*)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| EDM2-L + DDO (downsampled*)Type=Ours, #Params=777M, NFE=63, Generation Resolution=512x512, Resizing=OpenCV INTER_LANCZOS4 to 256x2562025.03 | 1.26 | 1.21 | |
| EDM2-L + DDO (downsampled*) + DPM-Solver-v3Type=Ours, #Params=777M, NFE=25, Generation Resolution=512x512, Resizing=OpenCV INTER_LANCZOS4 to 256x2562025.03 | 1.29 | 1.21 | |
| VAR-d30 (w/o sampling tricks) + DDOType=Ours, #Params=2.0B, NFE=10, sampling=no tricks2025.03 | 1.79 | 1.73 | |
| LightningDiTType=Diffusion, #Params=675M, NFE=2502025.03 | 2.17 | 1.35 | |
| VAR-d30Type=AR, #Params=2.0B, NFE=10, sampling=with top-k/top-p tricks2025.03 | 2.17 | 1.9 | |
| StyleGAN-XLType=GAN, #Params=166M, NFE=12025.03 | 2.3 | — | |
| MAR-HType=Masked, #Params=943M, NFE=2562025.03 | 2.35 | 1.55 | |
| VDM++Type=Diffusion, #Params=2B, NFE=5122025.03 | 2.4 | 2.12 | |
| U-ViT, LType=Diffusion, #Params=2B, NFE=5122025.03 | 2.77 | 2.44 | |
| VAR-d16 (w/o sampling tricks) + DDOType=Ours, #Params=310M, NFE=10, sampling=no tricks2025.03 | 3.12 | 2.54 | |
| VAR-d30 (w/o sampling tricks) + CCAType=Ours, #Params=2.0B, NFE=10, sampling=no tricks2025.03 | 3.3 | — | |
| RINType=Diffusion, #Params=410M, NFE=10002025.03 | 3.42 | — | |
| VAR-d16Type=AR, #Params=310M, NFE=10, sampling=with top-k/top-p tricks2025.03 | 3.62 | 3.3 | |
| MAGVIT-v2Type=Masked, #Params=307M, NFE=642025.03 | 3.65 | 1.78 | |
| RAR-XXLType=AR, #Params=1.5B, NFE=2562025.03 | 3.91 | 1.48 | |
| ViT-VQGANType=AR, #Params=1.7B, NFE=10242025.03 | 4.17 | — | |
| VAR-d30 (w/o sampling tricks)Type=Ours, #Params=2.0B, NFE=10, sampling=no tricks2025.03 | 4.74 | 1.92 | |
| SiT-XL + REPAType=Diffusion, #Params=675M, NFE=2502025.03 | 5.9 | 1.42 | |
| MaskGITType=Masked, #Params=227M, NFE=82025.03 | 6.18 | — | |
| BigGAN-deepType=GAN, #Params=112M, NFE=12025.03 | 6.95 | — | |
| ADM-UType=Diffusion, #Params=608M, NFE=5002025.03 | 7.49 | 3.94 | |
| RQ-TransformerType=AR, #Params=3.8B, NFE=682025.03 | 7.55 | — | |
| SiT-XLType=Diffusion, #Params=675M, NFE=2502025.03 | 8.3 | 2.06 | |
| DiT-XL/2Type=Diffusion, #Params=675M, NFE=2502025.03 | 9.62 | 2.27 | |
| Open-MAGVIT2-XLType=AR, #Params=1.5B, NFE=2562025.03 | 9.63 | 2.33 | |
| LDM-4Type=Diffusion, #Params=400M, NFE=2502025.03 | 10.56 | 3.6 | |
| ADMType=Diffusion, #Params=554M, NFE=2502025.03 | 10.94 | 4.59 | |
| VAR-d16 (w/o sampling tricks)Type=Ours, #Params=310M, NFE=10, sampling=no tricks2025.03 | 11.33 | 3.71 | |
| LlamaGen-3BType=AR, #Params=3.1B, NFE=2562025.03 | 13.58 | 3.05 | |
| VQGANType=AR, #Params=1.4B, NFE=2562025.03 | 15.78 | — | |
| GigaGANType=GAN, #Params=569M, NFE=12025.03 | — | 3.45 | |
| MaskBitType=Masked, #Params=305M, NFE=2562025.03 | — | 1.52 | |
| xAR-HType=AR, #Params=1.1B, NFE=502025.03 | — | 1.24 |