Class-conditional Image Generation on ImageNet 512x512
1.73FIDMAR-L, Diff Loss
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| MAR-L, Diff LossToken Type=continuous-valued tokens, #params=481M, CFG=true2024.06 | 1.73 | 279.9 | — | — | — | — | — | — | |
| MAR-LType=Mask, #Para.=481M2025.07 | 1.73 | 279.9 | — | — | — | — | — | — | |
| SiT-XL/2 + IGEpochs=602025.12 | 1.78 | 286.3 | 0.81 | 0.62 | — | 3.92 | — | — | |
| EDM2-XXLToken Type=continuous-valued tokens, #params=1.5B, CFG=true2024.06 | 1.81 | — | — | — | — | — | — | — | |
| MAGVIT-v2Token Type=vector-quantized tokens, #params=307M, CFG=true2024.06 | 1.91 | 324.3 | — | — | — | — | — | — | |
| EDM2-XXLToken Type=continuous-valued tokens, #params=1.5B, CFG=false2024.06 | 1.91 | — | — | — | — | — | — | — | |
| MAGVIT-v2Type=Mask, #Para.=307M, #Steps=642025.07 | 1.91 | 324.3 | — | — | — | — | — | — | |
| FreqFlow-L#params.=507M2026.04 | 2.02 | 285.3 | — | — | — | — | — | — | |
| SiT-XL/2 + REPAEpochs=2002025.12 | 2.08 | 274.6 | 0.83 | 0.58 | — | 4.19 | — | — | |
| Raster Counterpart-XLType=AR, #Para.=752M, #Steps=1024, Latency(s)=20.93, Throughput(img/s)=2.362025.07 | 2.09 | 315 | 81 | 57 | — | — | — | — | |
| LPD-XLType=Parallelized AR, #Para.=752M, #Steps=48, Latency(s)=1.01, Throughput(img/s)=18.182025.07 | 2.1 | 326 | 80 | 63 | — | — | — | — | |
| TiTok-BTok. Model=VQ, # Tokens=128, Classifier-Free Guidance (CFG)=true2024.12 | 2.13 | 261.2 | — | — | 1.52 | — | — | — | |
| SiT-XLTok. Model=SoftVQ-BL, # Tokens=64, Gflops=29.23, Throughput (imgs/sec)=5.38, Classifier-Free Guidance (CFG)=true2024.12 | 2.21 | 290.5 | — | — | 0.71 | — | — | — | |
| UVIT-2B+Tok. Model=AE, # Tokens=64, Gflops=9.44, Classifier-Free Guidance (CFG)=true2024.12 | 2.25 | — | — | — | 104.18 | — | — | — | |
| SiT-XL/2 + REPAEpochs=1002025.12 | 2.32 | 255.7 | 0.84 | 0.56 | — | 4.16 | — | — | |
| StyleGAN-XLArchitecture Type=GAN2022.09 | 2.41 | — | — | — | — | — | — | — | |
| StyleGAN-XLClassifier-Free Guidance (CFG)=true2024.12 | 2.41 | 267.7 | — | — | — | — | — | — | |
| StyleGAN-XLParams(M)=1662026.03 | 2.41 | — | 77 | 52 | — | — | — | — | |
| StyleGAN-XL2026.04 | 2.41 | 267.8 | — | — | — | — | — | — | |
| SiT-XL/2 + REPAEpochs=802025.12 | 2.44 | 247.3 | 0.84 | 0.56 | — | 4.21 | — | — | |
| MPDiT-XLParams(M)=482, GFLOPs=228.42026.03 | 2.47 | — | 83 | 56 | — | — | — | — | |
| MaskDiTEpochs=8002025.12 | 2.5 | 256.3 | 0.83 | 0.56 | — | 5.1 | — | — | |
| ComposerBackbone=VAR d-362026.03 | 2.51 | 305.7 | — | — | — | — | 1.07 | 21.96 | |
| UVIT-H+Tok. Model=AE, # Tokens=256, Gflops=6.14, Classifier-Free Guidance (CFG)=true2024.12 | 2.53 | — | — | — | 128.9 | — | — | — | |
| DiCo-XLParams(M)=701, GFLOPs=349.82026.03 | 2.53 | — | 83 | 56 | — | — | — | — | |
| Raster Counterpart-LType=AR, #Para.=337M, #Steps=1024, Latency(s)=14.25, Throughput(img/s)=3.792025.07 | 2.54 | 278.5 | 80 | 58 | — | — | — | — | |
| LPD-LType=Parallelized AR, #Para.=337M, #Steps=48, Latency(s)=0.69, Throughput(img/s)=35.162025.07 | 2.54 | 292.2 | 81 | 55 | — | — | — | — | |
| Test-time TrainingBackbone=VAR d-362026.03 | 2.54 | 305.4 | — | — | — | — | 112.37 | 32.45 | |
| SiT-XL/2Tok. Model=KL, # Tokens=4096, Gflops=0.10, Classifier-Free Guidance (CFG)=true2024.12 | 2.62 | 252.2 | — | — | 373.32 | — | — | — | |
| SiT-XL/2Epochs=6002025.12 | 2.62 | 252.2 | 0.84 | 0.57 | — | 4.18 | — | — | |
| SiT-XL/2Type=Diffusion, #Para.=675M, #Steps=2502025.07 | 2.62 | 252.21 | 84 | 57 | — | — | — | — | |
| SiT-XLParams(M)=675, GFLOPs=524.72026.03 | 2.62 | — | 84 | 57 | — | — | — | — | |
| VAR-d36-sType=VAR, Epoch=350, #Params=2.35B, Gflops=-2024.06 | 2.63 | 303.2 | — | — | — | — | — | — | |
| VAR-d36-sType=VAR, #Para.=2.3B, #Steps=10, Latency(s)=0.45, Throughput(img/s)=OOM2025.07 | 2.63 | 303.2 | — | — | — | — | — | — | |
| VAR-d36-sParams(M)=23002026.03 | 2.63 | — | — | — | — | — | — | — | |
| StandardBackbone=VAR d-362026.03 | 2.63 | 303.2 | — | — | — | — | 1 | 21.56 | |
| VAR-d36-s#params.=2.35B2026.04 | 2.63 | 303.2 | — | — | — | — | — | — | |
| VAR-d36-sEpoch=3502026.07 | 2.63 | 303.2 | — | — | — | — | — | — | |
| VDM++Token Type=pixel-based, #params=2B, CFG=true2024.06 | 2.65 | 278.1 | — | — | — | — | — | — | |
| VDM++2025.12 | 2.65 | 278.1 | — | — | — | — | — | — | |
| VDM++Params(M)=20002026.03 | 2.65 | — | — | — | — | — | — | — | |
| UVIT-H+Tok. Model=AE, # Tokens=64, Gflops=15.23, Classifier-Free Guidance (CFG)=true2024.12 | 2.66 | — | — | — | 32.99 | — | — | — | |
| DiffiTToken Type=continuous-valued tokens, CFG=true2024.06 | 2.67 | 252.1 | — | — | — | — | — | — | |
| DiffiTType=Diff., Epoch=800, #Params=561M, Gflops=-2024.06 | 2.67 | 252.1 | 83 | 55 | — | — | — | — | |
| DiffiT#params.=561M2026.04 | 2.67 | 252.1 | — | — | — | — | — | — | |
| MAR-L, Diff LossToken Type=continuous-valued tokens, #params=481M, CFG=false2024.06 | 2.74 | 205.2 | — | — | — | — | — | — | |
| ComposerBackbone=DiT-XL/22026.03 | 2.81 | 248.1 | — | — | — | — | 81.03 | 8.5 | |
| DiT-XLTok. Model=AE, # Tokens=256, Gflops=1.02, Classifier-Free Guidance (CFG)=true2024.12 | 2.84 | — | — | — | 80.75 | — | — | — | |
| ADM-G, ADM-UEpochs=4002025.12 | 2.85 | 221.7 | 0.84 | 0.53 | — | 5.86 | — | — | |
| Test-time TrainingBackbone=DiT-XL/22026.03 | 2.87 | 245.2 | — | — | — | — | 152.24 | 14.3 | |
| DiMR-XL/3RType=Diff., Epoch=800, #Params=525M, Gflops=2062024.06 | 2.89 | 289.8 | 83 | 55 | — | — | — | — | |
| DiMR-XL/3R#params.=525M2026.04 | 2.89 | 289.8 | — | — | — | — | — | — | |
| VDM++Token Type=pixel-based, #params=2B, CFG=false2024.06 | 2.99 | 232.2 | — | — | — | — | — | — | |
| PPF-XL-2Size=512, Training steps=400k, Training FLOPs (%)=7.6, Testing FLOPs (%)=58.72025.06 | 3.01 | 249.98 | 84 | 54 | — | 5.24 | — | — | |
| DiT-XL/2Token Type=continuous-valued tokens, #params=675M, CFG=true2024.06 | 3.04 | 240.8 | — | — | — | — | — | — | |
| DiT-XL/2Type=Diff., Epoch=599, #Params=675M, Gflops=5252024.06 | 3.04 | 240.8 | 84 | 54 | — | — | — | — | |
| DiT-XL/2Tok. Model=KL, # Tokens=4096, Gflops=0.10, Classifier-Free Guidance (CFG)=true2024.12 | 3.04 | 240.8 | — | — | 373.34 | — | — | — | |
| DiT-XL/2Epochs=6002025.12 | 3.04 | 240.8 | 0.84 | 0.54 | — | 5.02 | — | — | |
| DiT-XL/2Size=512, Training steps=-, Training FLOPs (%)=100, Testing FLOPs (%)=1002025.06 | 3.04 | 240.82 | 84 | 54 | — | 5.02 | — | — | |
| DiT-XL/2Type=Diffusion, #Para.=675M, #Steps=250, Latency(s)=11.32, Throughput(img/s)=0.102025.07 | 3.04 | 240.82 | 84 | 54 | — | — | — | — | |
| DiT-XL/2Params(M)=675, GFLOPs=524.72026.03 | 3.04 | — | 84 | 54 | — | — | — | — | |
| StandardBackbone=DiT-XL/22026.03 | 3.04 | 240.8 | — | — | — | — | 81 | 8.3 | |
| DiT-XL/2#params.=675M2026.04 | 3.04 | 240.8 | — | — | — | — | — | — | |
| DiT-XL/22026.07 | 3.04 | 240.8 | — | — | — | — | — | — | |
| PPF-XL-3Size=512, Training steps=400k, Training FLOPs (%)=5.8, Testing FLOPs (%)=45.42025.06 | 3.06 | 249.91 | 83 | 53 | — | 5.31 | — | — | |
| MAGVIT-v2Token Type=vector-quantized tokens, #params=307M, CFG=false2024.06 | 3.07 | 213.1 | — | — | — | — | — | — | |
| DiMR-XL/3RType=Diff., Epoch=400, #Params=525M, Gflops=2062024.06 | 3.23 | 285.1 | 82 | 54 | — | — | — | — | |
| DiT-XL/2 [33]*NFE=50, MACs=26.3T, Speed=1×2026.03 | 3.33 | 203.5 | 83.42 | 54.5 | — | 4.53 | — | — | |
| ARPG-XLType=Parallelized AR, #Para.=719M2025.07 | 3.38 | 257.8 | — | — | — | — | — | — | |
| DIFFUSSM-XLType=Diff., Epoch=236, #Params=673M, Gflops=10662024.06 | 3.41 | 255 | 85 | 49 | — | — | — | — | |
| DiffuSSM-XLParams(M)=673, GFLOPs=1066.22026.03 | 3.41 | — | 85 | 49 | — | — | — | — | |
| DIFFUSSM-XL#params.=673M2026.04 | 3.41 | 255 | — | — | — | — | — | — | |
| MAR-HTok. Model=SoftVQ-BL, # Tokens=64, Gflops=86.55, Throughput (imgs/sec)=1.50, Classifier-Free Guidance (CFG)=true2024.12 | 3.42 | 261.8 | — | — | 0.71 | — | — | — | |
| Timestep-Aware Block Masking (Ours)NFE=50, Extra Data=✗, Training Time=2.8h, MACs=17.9T, Speed=1.48×2026.03 | 3.64 | 202.8 | 83.18 | 54.5 | — | 5.06 | — | — | |
| L2C [29]*NFE=50, Extra Data=✓, Training Time=18.4h, MACs=16.9T, Speed=1.54×2026.03 | 3.76 | 200.7 | 83.16 | 54.3 | — | 5.1 | — | — | |
| DiM-HParams(M)=860, GFLOPs=7082026.03 | 3.78 | — | — | — | — | — | — | — | |
| ADM-G, ADM-UArchitecture Type=Diffusion based on U-Net, #Params=422M + 43M (Cls) + 309M (SR)2022.09 | 3.85 | — | — | — | — | — | — | — | |
| ADM-G, ADM-UType=Diff., Epoch=1081, #Params=731M, Gflops=28132024.06 | 3.85 | 221.7 | 84 | 53 | — | — | — | — | |
| ADMClassifier-Free Guidance (CFG)=true2024.12 | 3.85 | 221.7 | — | — | — | — | — | — | |
| ADM-UParams(M)=731, GFLOPs=28132026.03 | 3.85 | — | 84 | 53 | — | — | — | — | |
| ADM-G, ADM-U#params.=731M2026.04 | 3.85 | 221.7 | — | — | — | — | — | — | |
| U-ViT-H/4Params(M)=501, GFLOPs=133.32026.03 | 4.03 | — | 84 | 48 | — | — | — | — | |
| GAT-XL/2Resolution=512×512, Epochs=152025.09 | 4.04 | — | — | — | — | — | — | — | |
| U-ViT-H/4Architecture Type=Diffusion based on ViT, #Params=501M + 84M (AE), Latent representations & Classifier-free guidance=true2022.09 | 4.05 | — | — | — | — | — | — | — | |
| U-ViT-H/2-GToken Type=continuous-valued tokens, #params=501M, CFG=true2024.06 | 4.05 | 263.8 | — | — | — | — | — | — | |
| U-ViT-H/4Type=Diff., Epoch=400, #Params=501M, Gflops=1332024.06 | 4.05 | 263.8 | 84 | 48 | — | — | — | — | |
| U-ViT-H/4Classifier-Free Guidance (CFG)=true, Sampler=DPM-Solver, Steps=502024.12 | 4.05 | 263.8 | — | — | 128.92 | — | — | — | |
| U-ViT-H/4#params.=501M2026.04 | 4.05 | 263.8 | — | — | — | — | — | — | |
| Simple diffusion (U-Net)Epochs=8002025.12 | 4.28 | 171 | — | — | — | — | — | — | |
| MaskGIT-reType=Mask., Epoch=300, #Params=227M, Gflops=-2024.06 | 4.46 | 342 | — | — | — | — | — | — | |
| MaskGIT-re#params.=227M2026.04 | 4.46 | 342 | — | — | — | — | — | — | |
| Simple diffusion (U-ViT, L)Epochs=8002025.12 | 4.53 | 205.3 | — | — | — | — | — | — | |
| Simple DiffParams(M)=20002026.03 | 4.53 | — | — | — | — | — | — | — | |
| U-ViT-L/4Type=Diff., Epoch=400, #Params=287M, Gflops=772024.06 | 4.67 | 213.3 | 87 | 45 | — | — | — | — | |
| U-ViT-L/4Params(M)=287, GFLOPs=76.52026.03 | 4.67 | — | 87 | 45 | — | — | — | — | |
| U-ViT-L/4#params.=287M2026.04 | 4.67 | 213.3 | — | — | — | — | — | — | |
| UVIT-2B+Tok. Model=AE, # Tokens=64, Gflops=9.44, Classifier-Free Guidance (CFG)=false2024.12 | 6.5 | — | — | — | 104.18 | — | — | — | |
| MEPA-d20Epoch=662026.07 | 7.29 | 199.2 | — | — | — | — | — | — | |
| MaskGITToken Type=vector-quantized tokens, #params=227M, CFG=false2024.06 | 7.32 | 156 | — | — | — | — | — | — | |
| MaskGITType=Mask., Epoch=300, #Params=227M, Gflops=-2024.06 | 7.32 | 156 | 78 | 50 | — | — | — | — |