Image Generation on ImageNet 256x256 (train val)
1.13FIDRAE
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| RAEDiffusion Space=Latent, params=839+415M, Gflops=146, Architecture/Backbone=DiTDH-XL/22026.01 | 1.13 | — | — | 262.6 | — | |
| DDT-XL/2Diffusion Space=Latent, params=675+49M, Gflops=119, Architecture/Backbone=DDT-XL/22026.01 | 1.26 | — | — | 310.6 | — | |
| k-DiffDiffusion Space=Latent, params=675+49M, Gflops=119, Architecture/Backbone=LightningDiT-XL/12026.01 | 1.34 | — | — | 301.7 | — | |
| LightningDiT-XL/1Diffusion Space=Latent, params=675+49M, Gflops=119, Architecture/Backbone=LightningDiT-XL/12026.01 | 1.35 | — | — | 295.3 | — | |
| REPADiffusion Space=Latent, params=675+49M, Gflops=119, Architecture/Backbone=SiT-XL/22026.01 | 1.42 | — | — | 305.7 | — | |
| SiD2Diffusion Space=Pixel, Gflops=137, Architecture/Backbone=UViT/22026.01 | 1.73 | — | — | — | — | |
| PixelFlowDiffusion Space=Pixel, params=677M, Gflops=2909, Architecture/Backbone=XL/42026.01 | 1.98 | — | — | 282.1 | — | |
| TVM-XL/2w (CFG scale)=1.75, NFE=4, # Params=678M2025.11 | 1.99 | — | — | — | — | |
| SiT-XL/2Diffusion Space=Latent, params=675+49M, Gflops=119, Architecture/Backbone=SiT-XL/22026.01 | 2.06 | — | — | 277.5 | — | |
| VDM++Diffusion Space=Pixel, params=2B, Gflops=555, Architecture/Backbone=UViT/22026.01 | 2.12 | — | — | 267.7 | — | |
| PixNerdDiffusion Space=Pixel, params=700M, Gflops=134, Architecture/Backbone=XL/162026.01 | 2.15 | — | — | 297 | — | |
| SiT-XL/2w (CFG scale)=1.5, NFE=250x2, # Params=675M2025.11 | 2.15 | — | — | — | — | |
| DiT-XL/2Diffusion Space=Latent, params=675+49M, Gflops=119, Architecture/Backbone=DiT-XL/22026.01 | 2.27 | — | — | 278.2 | — | |
| DiT-XL/2w (CFG scale)=1.5, NFE=250x2, # Params=675M2025.11 | 2.27 | — | — | — | — | |
| SiDDiffusion Space=Pixel, params=2B, Gflops=555, Architecture/Backbone=UViT/22026.01 | 2.44 | — | — | 256.3 | — | |
| IMM-XL/2NFE=2x4, # Params=675M2025.11 | 2.51 | — | — | — | — | |
| TVM-XL/2w (CFG scale)=1.75, NFE=2, # Params=678M2025.11 | 2.61 | — | — | — | — | |
| TVM-XL/2w (CFG scale)=2, NFE=2, # Params=678M2025.11 | 2.8 | — | — | — | — | |
| MeanFlow-XL/2NFE=2, # Params=676M2025.11 | 2.93 | — | — | — | — | |
| DiT-XL/2w (CFG scale)=1.25, NFE=250x2, # Params=675M2025.11 | 3.22 | — | — | — | — | |
| TVM-XL/2w (CFG scale)=2, NFE=1, # Params=678M2025.11 | 3.29 | — | — | — | — | |
| RINDiffusion Space=Pixel, params=410M, Gflops=334, Architecture/Backbone=RIN2026.01 | 3.42 | — | — | 182 | — | |
| MeanFlow-XL/2NFE=1, # Params=676M2025.11 | 3.43 | — | — | — | — | |
| LDM-4-GNFE=250x2, # Params=400M2025.11 | 3.6 | — | — | — | — | |
| k-DiffDiffusion Space=Pixel, params=131M, Gflops=25, Architecture/Backbone=JiT-B/162026.01 | 3.64 | — | — | 274.9 | — | |
| JiT-B/16Diffusion Space=Pixel, params=131M, Gflops=25, Architecture/Backbone=JiT-B/162026.01 | 3.66 | — | — | 275.1 | — | |
| IMM-XL/2NFE=2x2, # Params=675M2025.11 | 3.99 | — | — | — | — | |
| TVM-XL/2w (CFG scale)=1.75, NFE=1, # Params=678M2025.11 | 4.58 | — | — | — | — | |
| ADM-GDiffusion Space=Pixel, params=554M, Gflops=1120, Architecture/Backbone=ADM-G2026.01 | 4.59 | — | — | 186.7 | — | |
| SSDD-XLspatial downsampling (f)=16, latent channels (c)=4, CFG=true, Sampling model=DiT-XL/22025.10 | 7.12 | — | — | — | 13.3 | |
| SSDD-Lspatial downsampling (f)=16, latent channels (c)=4, CFG=true, Sampling model=DiT-XL/22025.10 | 7.34 | — | — | — | 13.55 | |
| SSDD-XLspatial downsampling (f)=8, latent channels (c)=4, CFG=true, Sampling model=DiT-XL/22025.10 | 7.46 | — | — | — | 3.5 | |
| SSDD-Lspatial downsampling (f)=8, latent channels (c)=4, CFG=true, Sampling model=DiT-XL/22025.10 | 7.48 | — | — | — | 3.52 | |
| KL-VAEspatial downsampling (f)=8, latent channels (c)=4, CFG=true, Sampling model=DiT-XL/22025.10 | 7.63 | — | — | — | 3.54 | |
| SSDD-Mspatial downsampling (f)=8, latent channels (c)=4, CFG=true, Sampling model=DiT-XL/22025.10 | 7.63 | — | — | — | 3.52 | |
| SSDD-Bspatial downsampling (f)=8, latent channels (c)=4, CFG=true, Sampling model=DiT-XL/22025.10 | 7.85 | — | — | — | 3.54 | |
| SSDD-Mspatial downsampling (f)=16, latent channels (c)=4, CFG=true, Sampling model=DiT-XL/22025.10 | 7.99 | — | — | — | 13.65 | |
| SSDD-Sspatial downsampling (f)=8, latent channels (c)=4, CFG=true, Sampling model=DiT-XL/22025.10 | 8.02 | — | — | — | 3.55 | |
| IMM-XL/2NFE=1x2, # Params=675M2025.11 | 8.05 | — | — | — | — | |
| SSDD-Bspatial downsampling (f)=16, latent channels (c)=4, CFG=true, Sampling model=DiT-XL/22025.10 | 8.98 | — | — | — | 13.89 | |
| SSDD-Sspatial downsampling (f)=16, latent channels (c)=4, CFG=true, Sampling model=DiT-XL/22025.10 | 9.83 | — | — | — | 13.98 | |
| Shortcut-XL/2NFE=1, # Params=675M2025.11 | 10.6 | — | — | — | — | |
| DiTo-XL-LPIPSspatial downsampling (f)=8, latent channels (c)=4, CFG=true, Sampling model=DiT-XL/22025.10 | 10.83 | — | — | — | 0.77 | |
| ADMNFE=250x2, # Params=554M2025.11 | 10.96 | — | — | — | — | |
| SSDD-XLspatial downsampling (f)=16, latent channels (c)=4, CFG=false, Sampling model=DiT-XL/22025.10 | 16.08 | — | — | — | 25.04 | |
| SSDD-Lspatial downsampling (f)=16, latent channels (c)=4, CFG=false, Sampling model=DiT-XL/22025.10 | 16.27 | — | — | — | 25.95 | |
| SSDD-XLspatial downsampling (f)=8, latent channels (c)=4, CFG=false, Sampling model=DiT-XL/22025.10 | 16.79 | — | — | — | 6.89 | |
| SSDD-Lspatial downsampling (f)=8, latent channels (c)=4, CFG=false, Sampling model=DiT-XL/22025.10 | 16.86 | — | — | — | 6.96 | |
| SSDD-Mspatial downsampling (f)=8, latent channels (c)=4, CFG=false, Sampling model=DiT-XL/22025.10 | 17.03 | — | — | — | 6.98 | |
| SSDD-Mspatial downsampling (f)=16, latent channels (c)=4, CFG=false, Sampling model=DiT-XL/22025.10 | 17.24 | — | — | — | 26.29 | |
| SSDD-Bspatial downsampling (f)=8, latent channels (c)=4, CFG=false, Sampling model=DiT-XL/22025.10 | 17.46 | — | — | — | 7.05 | |
| SSDD-Sspatial downsampling (f)=8, latent channels (c)=4, CFG=false, Sampling model=DiT-XL/22025.10 | 17.77 | — | — | — | 7.07 | |
| SSDD-Bspatial downsampling (f)=16, latent channels (c)=4, CFG=false, Sampling model=DiT-XL/22025.10 | 18.89 | — | — | — | 27.2 | |
| SSDD-XLspatial downsampling (f)=16, latent channels (c)=16, CFG=true, Sampling model=DiT-XL/22025.10 | 19.68 | — | — | — | 13.31 | |
| KL-VAEspatial downsampling (f)=8, latent channels (c)=4, CFG=false, Sampling model=DiT-XL/22025.10 | 19.71 | — | — | — | 7.05 | |
| SSDD-Lspatial downsampling (f)=16, latent channels (c)=16, CFG=true, Sampling model=DiT-XL/22025.10 | 19.74 | — | — | — | 13.57 | |
| SSDD-Mspatial downsampling (f)=16, latent channels (c)=16, CFG=true, Sampling model=DiT-XL/22025.10 | 19.91 | — | — | — | 13.66 | |
| SSDD-Sspatial downsampling (f)=16, latent channels (c)=4, CFG=false, Sampling model=DiT-XL/22025.10 | 20.28 | — | — | — | 27.55 | |
| SSDD-Bspatial downsampling (f)=16, latent channels (c)=16, CFG=true, Sampling model=DiT-XL/22025.10 | 20.28 | — | — | — | 13.9 | |
| iCT-XL/2NFE=2, # Params=675M2025.11 | 20.3 | — | — | — | — | |
| SSDD-Sspatial downsampling (f)=16, latent channels (c)=16, CFG=true, Sampling model=DiT-XL/22025.10 | 20.51 | — | — | — | 13.98 | |
| CoReDi-L/16Iter.=200K, Param.=426M, Guidance=None2026.04 | 21.5 | — | — | — | — | |
| DiTo-XL-LPIPSspatial downsampling (f)=8, latent channels (c)=4, CFG=false, Sampling model=DiT-XL/22025.10 | 22.47 | — | — | — | 0.87 | |
| SSDD-XLspatial downsampling (f)=16, latent channels (c)=16, CFG=false, Sampling model=DiT-XL/22025.10 | 25.01 | — | — | — | 25.07 | |
| SSDD-Lspatial downsampling (f)=16, latent channels (c)=16, CFG=false, Sampling model=DiT-XL/22025.10 | 25.03 | — | — | — | 26.02 | |
| SSDD-Mspatial downsampling (f)=16, latent channels (c)=16, CFG=false, Sampling model=DiT-XL/22025.10 | 25.2 | — | — | — | 26.34 | |
| SSDD-Bspatial downsampling (f)=16, latent channels (c)=16, CFG=false, Sampling model=DiT-XL/22025.10 | 25.59 | — | — | — | 27.24 | |
| SSDD-Sspatial downsampling (f)=16, latent channels (c)=16, CFG=false, Sampling model=DiT-XL/22025.10 | 25.84 | — | — | — | 27.57 | |
| KL-VAEspatial downsampling (f)=16, latent channels (c)=16, CFG=true, Sampling model=DiT-XL/22025.10 | 26.42 | — | — | — | 14.04 | |
| DeCo-L/16Iter.=200K, Param.=426M, Guidance=None2026.04 | 31.3 | — | — | — | — | |
| CoReDi-L/16Iter.=100K, Param.=426M, Guidance=None2026.04 | 31.5 | — | — | — | — | |
| KL-VAEspatial downsampling (f)=16, latent channels (c)=16, CFG=false, Sampling model=DiT-XL/22025.10 | 31.96 | — | — | — | 27.79 | |
| iCT-XL/2NFE=1, # Params=675M2025.11 | 34.24 | — | — | — | — | |
| DeCo-L/16Iter.=100K, Param.=426M, Guidance=None, Note=Reproduction2026.04 | 46 | — | — | — | — | |
| BigGANImplementation=Reproduced by authors, G Params=82.10M2020.11 | — | 9.95 | 9.88 | 187.6 | — | |
| Omni-GANG Params=82.10M2020.11 | — | 6.03 | 6.83 | 304.05 | — | |
| Omni-INR-GANG Params=82.19M2020.11 | — | 4.93 | 6.36 | 343.22 | — |