Class-conditional Image Generation on ImageNet 64x64
1.03FIDEDM2-S-autog + C2FG
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| EDM2-S-autog + C2FGAutoguidance=true, Guidance Scale (omega_0)=1.7, Guidance Scale (lambda)=0.05, Evaluation Space=Pixel Space2026.03 | 1.03 | — | — | — | — | — | — | |
| EDM2-S-autogAutoguidance=true, Guidance Scale (omega)=1.7, Evaluation Space=Pixel Space2026.03 | 1.04 | — | — | — | — | — | — | |
| RIN# Fwd Pass=10002024.05 | 1.23 | — | — | — | — | — | — | |
| RIN# param=281M, NFE=10002024.06 | 1.23 | — | — | 67 | — | — | — | |
| Moment Matching Alternating# param=400M, NFE=82024.06 | 1.24 | — | — | 78 | — | — | — | |
| DMD2# Fwd Pass=1, training=longer training2024.05 | 1.28 | — | — | — | — | — | — | |
| Moment Matching Instant# param=400M, NFE=82024.06 | 1.35 | — | — | 81 | — | — | — | |
| EDMArchitecture Type=Diffusion based on U-Net, #Params=296M, Advanced training techniques=true2022.09 | 1.36 | — | — | — | — | — | — | |
| EDM# Fwd Pass=511, mode=Teacher, SDE2024.05 | 1.36 | — | — | — | — | — | — | |
| EDMModel category=Diffusion Models / Flow Matching2025.11 | 1.36 | — | — | — | — | — | — | |
| EDM teacher (SDE)FP=5112026.05 | 1.36 | — | — | — | — | — | — | |
| MultiStep-CD (Heek et al., 2024)# param=1.2B, NFE=82024.06 | 1.4 | — | — | 73 | — | — | — | |
| our base model# param=400M, NFE=10242024.06 | 1.42 | — | — | 84 | — | — | — | |
| VDM++# param=2B, NFE=10242024.06 | 1.43 | — | — | — | — | — | — | |
| TeacherNFE=2562024.05 | 1.43 | — | — | — | — | — | — | |
| CDMArchitecture Type=Diffusion based on U-Net, #Params=Unknown2022.09 | 1.48 | — | — | — | — | — | — | |
| CDM2023.03 | 1.48 | — | — | — | — | — | — | |
| Moment Matching Alternating# param=400M, NFE=42024.06 | 1.5 | — | — | 75 | — | — | — | |
| StyleGAN-XLArchitecture Type=GAN2022.09 | 1.51 | — | — | — | — | — | — | |
| StyleGAN-XL2023.03 | 1.51 | — | — | — | — | — | — | |
| DMD2# Fwd Pass=12024.05 | 1.51 | — | — | — | — | — | — | |
| DMD2 (200K)FP=12026.05 | 1.51 | — | — | — | — | — | — | |
| StyleGAN-XL# Fwd Pass=12024.05 | 1.52 | — | — | — | — | — | — | |
| EDM2-SAutoguidance=false, Evaluation Space=Pixel Space2026.03 | 1.58 | — | — | — | — | — | — | |
| TFD (Ours) (200K)FP=12026.05 | 1.58 | — | — | — | — | — | — | |
| MultiStep-CD (Heek et al., 2024)# param=1.2B, NFE=42024.06 | 1.6 | — | — | — | — | — | — | |
| PD (Salimans & Ho, 2022)# param=400M, NFE=82024.06 | 1.7 | — | — | 63 | — | — | — | |
| CTM (Kim et al., 2024)NFE=22024.06 | 1.73 | — | — | 64 | — | — | — | |
| MultiStep-CD (Heek et al., 2024)# param=1.2B, NFE=22024.06 | 1.9 | — | — | — | — | — | — | |
| CTM# Fwd Pass=12024.05 | 1.92 | — | — | — | — | — | — | |
| CTMModel Type=Diffusion Models, NFE=12024.12 | 1.92 | — | — | — | — | — | — | |
| FreqFlow-B#params.=134M2026.04 | 1.92 | — | — | 59.34 | — | — | — | |
| CTMFP=12026.05 | 1.92 | — | — | — | — | — | — | |
| Multistep CDNFE=2, Sampling=Multiple Steps2024.05 | 2 | — | — | — | — | — | — | |
| ADMArchitecture Type=Diffusion based on U-Net, #Params=296M2022.09 | 2.07 | — | — | — | — | — | — | |
| ADM# Fwd Pass=2502024.05 | 2.07 | — | — | — | — | — | — | |
| ADM# Fwd Pass=2502023.11 | 2.07 | — | — | — | — | — | — | |
| ADMModel category=Diffusion Models / Flow Matching2025.11 | 2.09 | — | — | — | — | — | 4.29 | |
| CCMModel Type=Flow Matching Models, NFE=12024.12 | 2.18 | — | — | — | — | — | — | |
| EMD-16 (ours)NFE=1, Sampling=Single Step2024.05 | 2.2 | — | — | — | — | 0.59 | — | |
| EMDFP=12026.05 | 2.2 | — | — | — | — | — | — | |
| DiMR-L/3RType=Diffusion, Epoch=240, Parameters=284M, Gflops=1112024.06 | 2.21 | — | — | 55.73 | 75 | 60 | — | |
| DiMR-L/3R#params.=284M2026.04 | 2.21 | — | — | 55.73 | — | — | — | |
| ViT-L (ours)#Params=269M2023.03 | 2.28 | — | — | — | — | — | — | |
| EDM# Fwd Pass=511, mode=Teacher, ODE2024.05 | 2.32 | — | — | — | — | — | — | |
| EDM (Teacher)# Fwd Pass=5122023.11 | 2.32 | — | — | — | — | — | — | |
| EDM teacher (ODE)FP=5112026.05 | 2.32 | — | — | — | — | — | — | |
| ADM-EST'=100, λ(t)=uniform, classifier guidance=none2023.08 | 2.39 | — | — | — | — | — | — | |
| PD (Salimans & Ho, 2022)# param=400M, NFE=42024.06 | 2.4 | — | — | — | — | — | — | |
| TRACTNFE=82024.06 | 2.41 | — | — | — | — | — | — | |
| EDMModel Type=Diffusion Models, NFE=792024.12 | 2.44 | — | — | — | — | — | — | |
| ADM#Params=296M2023.03 | 2.61 | — | — | — | — | — | — | |
| DMD# Fwd Pass=12024.05 | 2.62 | — | — | — | — | — | — | |
| DMD (Yin et al., 2023)NFE=12024.06 | 2.62 | — | — | — | — | — | — | |
| DMD# Fwd Pass=12023.11 | 2.62 | — | — | — | — | — | — | |
| DMDNFE=1, Sampling=Single Step2024.05 | 2.62 | — | — | — | — | — | — | |
| DMDFP=12026.05 | 2.62 | — | — | — | — | — | — | |
| ADMT'=100, classifier guidance=none2023.08 | 2.71 | — | — | — | — | — | — | |
| iCT-deepNFE=22024.06 | 2.77 | — | — | — | — | — | — | |
| IDDPM (large)Architecture Type=Diffusion based on U-Net, #Params=270M2022.09 | 2.92 | — | — | — | — | — | — | |
| IDDPM (large)#Params=270M2023.03 | 2.92 | — | — | — | — | — | — | |
| iDDPMModel category=Diffusion Models / Flow Matching2025.11 | 2.92 | — | — | — | — | — | 3.79 | |
| TRACTNFE=42024.06 | 2.93 | — | — | — | — | — | — | |
| Moment Matching Alternating# param=400M, NFE=12024.06 | 3 | — | — | 89 | — | — | — | |
| ADM-EST'=50, λ(t)=uniform, classifier guidance=none2023.08 | 3.07 | — | — | — | — | — | — | |
| EMD-1 (baseline)NFE=1, Sampling=Single Step2024.05 | 3.1 | — | — | — | — | 0.55 | — | |
| MultiStep-CD (Heek et al., 2024)# param=1.2B, NFE=12024.06 | 3.2 | — | — | — | — | — | — | |
| iCT-deep# Fwd Pass=12024.05 | 3.25 | — | — | — | — | — | — | |
| iCT-deepNFE=12024.06 | 3.25 | — | — | — | — | — | — | |
| iCT-deepModel category=Consistency Models (CMs)2025.11 | 3.25 | — | — | — | — | — | — | |
| iCT-deepFP=12026.05 | 3.25 | — | — | — | — | — | — | |
| Moment Matching Instant# param=400M, NFE=42024.06 | 3.4 | — | — | 98 | — | — | — | |
| DDPMImplementation=our repo., Data Encoding=continuous pixels2022.08 | 3.43 | — | — | — | — | — | — | |
| DiMR-M/3RType=Diffusion, Epoch=240, Parameters=133M, Gflops=542024.06 | 3.65 | — | — | 42.41 | 74 | 59 | — | |
| DiMR-M/3R#params.=133M2026.04 | 3.65 | — | — | 42.41 | — | — | — | |
| TARFlow + R-REPAModel category=Normalizing Flows2025.11 | 3.69 | — | — | — | — | — | 4.34 | |
| ADMT'=50, classifier guidance=none2023.08 | 3.75 | — | — | — | — | — | — | |
| Moment Matching Alternating# param=400M, NFE=22024.06 | 3.86 | — | — | 60 | — | — | — | |
| iCTModel Type=Diffusion Models, NFE=12024.12 | 4.02 | — | — | — | — | — | — | |
| ECMModel Type=Diffusion Models, NFE=12024.12 | 4.05 | — | — | — | — | — | — | |
| BigGAN-deepArchitecture Type=GAN2022.09 | 4.06 | — | — | — | — | — | — | |
| BigGAN-deep2023.03 | 4.06 | — | — | — | — | — | — | |
| BigGAN-deep# Fwd Pass=12024.05 | 4.06 | — | — | — | — | — | — | |
| BigGAN-deep# Fwd Pass=12023.11 | 4.06 | — | — | — | — | — | — | |
| BigGAN-deepNFE=1, Sampling=Single Step2024.05 | 4.06 | — | — | — | — | 0.48 | — | |
| BigGANModel category=Generative Adversarial Networks (GANs)2025.11 | 4.06 | — | — | — | — | — | 3.96 | |
| TARFlowModel category=Normalizing Flows2025.11 | 4.21 | — | — | — | — | — | 5.34 | |
| U-ViT-L/4Architecture Type=Diffusion based on ViT, #Params=287M2022.09 | 4.26 | — | — | — | — | — | — | |
| U-ViT-Large#Params=287M2023.03 | 4.26 | — | — | — | — | — | — | |
| U-ViT-L/4Type=Diffusion, Epoch=240, Parameters=287M, Gflops=772024.06 | 4.26 | — | — | 40.66 | 71 | 62 | — | |
| U-ViT-L/4#params.=287M2026.04 | 4.26 | — | — | 40.66 | — | — | — | |
| CD (LPIPS)NFE=32024.06 | 4.32 | — | — | — | — | — | — | |
| CD (LPIPS)NFE=22024.06 | 4.7 | — | — | — | — | — | — | |
| PD (Salimans & Ho, 2022)# param=400M, NFE=22024.06 | 4.7 | — | — | — | — | — | — | |
| CDNFE=2, Sampling=Multiple Steps2024.05 | 4.7 | — | — | — | — | 0.64 | — | |
| CD (LPIPS)Model category=Consistency Models (CMs)2025.11 | 4.7 | — | — | — | — | — | — | |
| Bit DiffusionData Encoding=UINT82022.08 | 4.84 | — | — | — | — | — | — | |
| INDISNFE=72026.03 | 4.94 | — | — | — | — | — | — | |
| INDISNFE=72026.03 | 4.94 | — | — | — | — | — | — | |
| TRACTNFE=22024.06 | 4.97 | — | — | — | — | — | — |