Image Synthesis on ImageNet 256x256
1.42FIDREPA
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| REPAArchitecture type=Latent diffusion, Transformer, Epochs=800, Classifier-Free Guidance (CFG)=true, CFG scheduling=interval from [30]2026.05 | 1.42 | 4.7 | 305.7 | 80 | 65 | |
| ERDArchitecture type=Latent diffusion, Transformer + U-Net hybrid, Epochs=800, Classifier-Free Guidance (CFG)=true, CFG scheduling=interval from [30]2026.05 | 1.45 | 4.51 | 315.4 | 81 | 62 | |
| Min-SNRArchitecture type=Latent diffusion, Transformer, Epochs=1400, Classifier-Free Guidance (CFG)=true, CFG scheduling=interval from [30]2026.05 | 1.57 | — | — | — | — | |
| MDTv2-XL/2Architecture type=Latent diffusion, Transformer + U-Net hybrid, Epochs=1080, Classifier-Free Guidance (CFG)=true, CFG scheduling=interval from [30]2026.05 | 1.58 | 4.52 | 314.7 | 79 | 65 | |
| DiffiTArchitecture type=Latent diffusion, Transformer + U-Net hybrid, Classifier-Free Guidance (CFG)=true, CFG scheduling=interval from [30]2026.05 | 1.73 | — | 276.5 | 80 | 62 | |
| ERDArchitecture type=Latent diffusion, Transformer + U-Net hybrid, Epochs=800, Classifier-Free Guidance (CFG)=true2026.05 | 1.89 | 4.48 | 282.5 | 82 | 60 | |
| SiT-XL/2Architecture type=Latent diffusion, Transformer, Epochs=1400, Classifier-Free Guidance (CFG)=true2026.05 | 2.06 | 4.5 | 270.3 | 82 | 59 | |
| DiT-XL/2Architecture type=Latent diffusion, Transformer, Epochs=1400, Classifier-Free Guidance (CFG)=true2026.05 | 2.27 | 4.6 | 278.2 | 83 | 57 | |
| MaskDiTArchitecture type=Latent diffusion, Transformer, Epochs=1600, Classifier-Free Guidance (CFG)=true2026.05 | 2.28 | 5.67 | 276.6 | 80 | 61 | |
| U-ViT-H/2Architecture type=Latent diffusion, Transformer + U-Net hybrid, Epochs=400, Classifier-Free Guidance (CFG)=true2026.05 | 2.29 | 5.68 | 263.9 | 82 | 57 | |
| VDM++Architecture type=Pixel diffusion, Epochs=560, Classifier-Free Guidance (CFG)=true2026.05 | 2.4 | — | 225.3 | — | — | |
| Simple diffusionArchitecture type=Pixel diffusion, Epochs=800, Classifier-Free Guidance (CFG)=true2026.05 | 2.77 | — | 211.8 | — | — | |
| SD-DiTArchitecture type=Latent diffusion, Transformer, Epochs=480, Classifier-Free Guidance (CFG)=true2026.05 | 3.23 | — | — | — | — | |
| LDM-4Architecture type=Latent diffusion, U-Net, Epochs=200, Classifier-Free Guidance (CFG)=true2026.05 | 3.6 | — | 247.7 | 87 | 48 | |
| ADM-UArchitecture type=Pixel diffusion, Epochs=400, Classifier-Free Guidance (CFG)=true2026.05 | 3.94 | 6.14 | 186.7 | 82 | 52 | |
| CDMArchitecture type=Pixel diffusion, Epochs=2160, Classifier-Free Guidance (CFG)=true2026.05 | 4.88 | — | 158.7 | — | — | |
| GADCData Budget=8.0% (100K), Model=SiT-L/2, Training Iterations=100K, Evaluation Samples=50K2026.06 | 8.83 | — | 167.5 | 0.74 | 0.53 | |
| D2CData Budget=8.0% (100K), Model=SiT-L/2, Training Iterations=100K, Evaluation Samples=50K2026.06 | 15.01 | — | 127.6 | 0.73 | 0.48 | |
| DQData Budget=8.0% (100K), Model=SiT-L/2, Training Iterations=100K, Evaluation Samples=50K2026.06 | 15.24 | — | 121.9 | 0.73 | 0.5 | |
| CCSData Budget=8.0% (100K), Model=SiT-L/2, Training Iterations=100K, Evaluation Samples=50K2026.06 | 15.41 | — | 142.9 | 0.74 | 0.43 | |
| Random†Data Budget=8.0% (100K), Model=SiT-L/2, Training Iterations=100K, Evaluation Samples=50K2026.06 | 15.58 | — | 125.9 | 0.73 | 0.5 |