Image Generation on ImageNet
1.52FIDMaskBit
Evaluation results
| Method | Links | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| MaskBitGen. Model=MaskBit, Type=Mask., #Token=256, #Step=2562026.02 | 1.52 | — | — | 328.6 | — | — | — | — | — | — | — | — | |
| CompTokGen. Model=ϵMaskGIT, Type=Mask., #Token=256, #Step=642026.02 | 1.6 | — | — | 297.9 | — | — | — | — | — | — | — | — | |
| MARGen. Model=MAR-L, Type=Mask., #Token=256, #Step=642026.02 | 1.78 | — | — | 296 | — | — | — | — | — | — | — | — | |
| FlexTokGen. Model=AR, Type=AR, #Token=32, #Step=322026.02 | 1.86 | — | — | — | — | — | — | — | — | — | — | — | |
| TiTok-S-128Gen. Model=MaskGIT-L, Type=Mask., #Token=128, #Step=642026.02 | 1.97 | — | — | 281.8 | — | — | — | — | — | — | — | — | |
| LuminarkBackbone Model=EDM2-XXL, Paradigm=Diffusion, Tokenizer=Continuous, Resolution=512×512, Param.=3.05 B2026.01 | 2.13 | — | — | — | — | — | — | — | — | — | — | — | |
| PF-ODE + MADMODE Solver=Heun, PC Sampler=PF-ODE + MADM2026.05 | 2.15 | — | — | — | — | — | — | — | — | — | — | — | |
| PF-ODE + MADMODE Solver=Euler, PC Sampler=PF-ODE + MADM2026.05 | 2.25 | — | — | — | — | — | — | — | — | — | — | — | |
| PF-ODEODE Solver=Heun, PC Sampler=PF-ODE2026.05 | 2.32 | — | — | — | — | — | — | — | — | — | — | — | |
| PF-ODEODE Solver=Euler, PC Sampler=PF-ODE2026.05 | 2.41 | — | — | — | — | — | — | — | — | — | — | — | |
| SemanticistGen. Model=ϵLlamaGen-L, Type=AR, #Token=32, #Step=322026.02 | 2.57 | — | — | 260.9 | — | — | — | — | — | — | — | — | |
| ImageFolderGen. Model=VAR-d16, Type=VAR, #Token=286, #Step=102026.02 | 2.6 | — | — | 295 | — | — | — | — | — | — | — | — | |
| LuminarkBackbone Model=EDM2-L, Paradigm=Diffusion, Tokenizer=Continuous, Resolution=512×512, Param.=1.56 B2026.01 | 2.72 | — | — | — | — | — | — | — | — | — | — | — | |
| CRTGen. Model=LlamaGen-L, Type=AR, #Token=256, #Step=2562026.02 | 2.75 | — | — | 265.2 | — | — | — | — | — | — | — | — | |
| TiTok-L-32Gen. Model=MaskGIT-L, Type=Mask., #Token=32, #Step=82026.02 | 2.77 | — | — | 194 | — | — | — | — | — | — | — | — | |
| VTP-L-d64Training Epochs=80, Classifier-Free Guidance (CFG)=false, Generation Model=LightningDiT, Model Scale=Large2025.12 | 2.81 | — | — | — | — | — | — | — | — | — | — | — | |
| LlamaGen-dist#Params=343M, Epochs=3002025.03 | 3.04 | — | — | 258.19 | — | — | — | — | — | — | — | — | |
| LuminarkBackbone Model=VAR-d30, Paradigm=AR., Tokenizer=Discrete, Resolution=256×256, Param.=2.01 B2026.01 | 3.06 | — | — | — | — | — | — | — | — | — | — | — | |
| LlamaGen-sft#Params=343M, Epochs=3002025.03 | 3.08 | — | — | 256.07 | — | — | — | — | — | — | — | — | |
| LuminarkBackbone Model=MAR-Huge, Paradigm=Hybrid, Tokenizer=Continuous, Resolution=256×256, Param.=0.94 B2026.01 | 3.09 | — | — | — | — | — | — | — | — | — | — | — | |
| VARGen. Model=VAR-d16, Type=VAR, #Token=680, #Step=102026.02 | 3.3 | — | — | 274.4 | — | — | — | — | — | — | — | — | |
| LuminarkBackbone Model=MAR-Large, Paradigm=Hybrid, Tokenizer=Continuous, Resolution=256×256, Param.=0.48 B2026.01 | 3.32 | — | — | — | — | — | — | — | — | — | — | — | |
| GigaGAN#Params=569M, Epochs=3002025.03 | 3.45 | — | — | 225.5 | — | — | — | — | — | — | — | — | |
| LDM-4#Params=400M, Epochs=3002025.03 | 3.6 | — | — | 247.7 | — | — | — | — | — | — | — | — | |
| RQ-VAEGen. Model=RQ-Trans., Type=AR, #Token=256, #Step=642026.02 | 3.8 | — | — | 323.7 | — | — | — | — | — | — | — | — | |
| LlamaGenGen. Model=LlamaGen-L, Type=AR, #Token=256, #Step=2562026.02 | 3.8 | — | — | 248.3 | — | — | — | — | — | — | — | — | |
| VFM-VAE#Images=44M, Epochs=642025.10 | 3.8 | — | — | 152.8 | — | — | — | — | — | — | — | — | |
| LuminarkBackbone Model=MAR-Base, Paradigm=Hybrid, Tokenizer=Continuous, Resolution=256×256, Param.=0.21 B2026.01 | 3.88 | — | — | — | — | — | — | — | — | — | — | — | |
| VTP-B-d64Training Epochs=80, Classifier-Free Guidance (CFG)=false, Generation Model=LightningDiT, Model Scale=Base2025.12 | 3.88 | — | — | — | — | — | — | — | — | — | — | — | |
| LuminarkBackbone Model=VAR-d16, Paradigm=AR., Tokenizer=Discrete, Resolution=256×256, Param.=0.31 B2026.01 | 4.04 | — | — | — | — | — | — | — | — | — | — | — | |
| Causal MARGen. Model=MAR-L, Type=AR, #Token=256, #Step=2562026.02 | 4.07 | — | — | 232.4 | — | — | — | — | — | — | — | — | |
| ViT-VQGANGen. Model=VIM-L, Type=AR, #Token=1024, #Step=10242026.02 | 4.17 | — | — | 175.1 | — | — | — | — | — | — | — | — | |
| LlamaGen-dist#Params=343M, Epochs=502025.03 | 4.18 | — | — | 209.41 | — | — | — | — | — | — | — | — | |
| LuminarkBackbone Model=VAR-d36, Paradigm=AR., Tokenizer=Discrete, Resolution=512×512, Param.=2.35 B2026.01 | 4.22 | — | — | — | — | — | — | — | — | — | — | — | |
| LlamaGen-sft#Params=343M, Epochs=502025.03 | 4.24 | — | — | 206.74 | — | — | — | — | — | — | — | — | |
| RAE-d768Training Epochs=80, Classifier-Free Guidance (CFG)=false, Generation Model=LightningDiT, Note=trained with same generation model2025.12 | 4.28 | — | — | — | — | — | — | — | — | — | — | — | |
| VA-VAE-d32Training Epochs=80, Classifier-Free Guidance (CFG)=false, Generation Model=LightningDiT, Latent Dimension=322025.12 | 4.29 | — | — | — | — | — | — | — | — | — | — | — | |
| LuminarkBackbone Model=EDM2-XS, Paradigm=Diffusion, Tokenizer=Continuous, Resolution=512×512, Param.=0.25 B2026.01 | 4.4 | — | — | — | — | — | — | — | — | — | — | — | |
| PRC-W.Backbone Model=EDM2-XXL, Paradigm=Diffusion, Tokenizer=Continuous, Resolution=512×512, Param.=3.05 B2026.01 | 4.6 | — | — | — | — | — | — | — | — | — | — | — | |
| GS.Backbone Model=EDM2-XXL, Paradigm=Diffusion, Tokenizer=Continuous, Resolution=512×512, Param.=3.05 B2026.01 | 4.87 | — | — | — | — | — | — | — | — | — | — | — | |
| PRC-W.Backbone Model=EDM2-L, Paradigm=Diffusion, Tokenizer=Continuous, Resolution=512×512, Param.=1.56 B2026.01 | 5 | — | — | — | — | — | — | — | — | — | — | — | |
| GS.Backbone Model=EDM2-L, Paradigm=Diffusion, Tokenizer=Continuous, Resolution=512×512, Param.=1.56 B2026.01 | 5.08 | — | — | — | — | — | — | — | — | — | — | — | |
| VA-VAE#Images=160M, Epochs=642025.10 | 5.14 | — | — | 130.2 | — | — | — | — | — | — | — | — | |
| VQGANGen. Model=Tam. Trans., Type=AR, #Token=256, #Step=2562026.02 | 5.2 | — | — | 280.3 | — | — | — | — | — | — | — | — | |
| Full-precision#Bits=16/16, Size (MB)=13492026.05 | 5.32 | — | — | 236.17 | — | — | — | — | — | — | — | — | |
| Q-DiT + NBC#Bits=8/8, Size (MB)=7072026.05 | 5.41 | — | — | 240.4 | — | — | — | — | — | — | — | — | |
| Q-DiT#Bits=8/8, Size (MB)=6832026.05 | 5.45 | — | — | 236.91 | — | — | — | — | — | — | — | — | |
| VTP-S-d64Training Epochs=80, Classifier-Free Guidance (CFG)=false, Generation Model=LightningDiT, Model Scale=Small2025.12 | 5.46 | — | — | — | — | — | — | — | — | — | — | — | |
| RepQ-ViT#Bits=8/8, Size (MB)=6772026.05 | 5.46 | — | — | 234.74 | — | — | — | — | — | — | — | — | |
| GPTQ#Bits=8/8, Size (MB)=6902026.05 | 5.9 | — | — | 218.9 | — | — | — | — | — | — | — | — | |
| Q-DiT + NBC#Bits=4/8, Size (MB)=3612026.05 | 6.12 | — | — | 215.7 | — | — | — | — | — | — | — | — | |
| LlamaGen-dist#Params=111M, Epochs=3002025.03 | 6.27 | — | — | 164.32 | — | — | — | — | — | — | — | — | |
| LlamaGen-sft#Params=111M, Epochs=3002025.03 | 6.44 | — | — | 157.17 | — | — | — | — | — | — | — | — | |
| Factorized Diffusion ArchitecturesResolution=642023.09 | 6.54 | — | — | — | — | — | — | — | — | — | — | — | |
| θ-TrapezoidalNFE=642026.07 | 6.59 | — | — | — | — | — | — | — | — | — | — | — | |
| Q-DiT#Bits=4/8, Size (MB)=3472026.05 | 6.75 | — | — | 208.38 | — | — | — | — | — | — | — | — | |
| Parallel (Ours)NFE=322026.07 | 6.82 | — | — | — | — | — | — | — | — | — | — | — | |
| PRC-W.Backbone Model=EDM2-XS, Paradigm=Diffusion, Tokenizer=Continuous, Resolution=512×512, Param.=0.25 B2026.01 | 6.88 | — | — | — | — | — | — | — | — | — | — | — | |
| GS.Backbone Model=EDM2-XS, Paradigm=Diffusion, Tokenizer=Continuous, Resolution=512×512, Param.=0.25 B2026.01 | 6.99 | — | — | — | — | — | — | — | — | — | — | — | |
| DDPMResolution=642023.09 | 7.02 | — | — | — | — | — | — | — | — | — | — | — | |
| SD-VAETraining Epochs=80, Classifier-Free Guidance (CFG)=false, Generation Model=LightningDiT2025.12 | 7.13 | — | — | — | — | — | — | — | — | — | — | — | |
| SD-VAE#Images=108M, Epochs=802025.10 | 7.13 | — | — | — | — | — | — | — | — | — | — | — | |
| τ-leapingNFE=642026.07 | 7.56 | — | — | — | — | — | — | — | — | — | — | — | |
| Sequential DDIMSolver=DDIM2024.12 | 9 | — | — | — | 1,024 | — | — | — | — | — | — | — | |
| SRDSSolver=DDIM, SRDS Iters=4.62024.12 | 9 | — | — | — | — | 175 | 4,612 | — | — | — | — | — | |
| GPTQ#Bits=4/8, Size (MB)=3512026.05 | 9.04 | — | — | 166.35 | — | — | — | — | — | — | — | — | |
| Uniform-radial VPTime-Conditioning=Uniform-radial VP2026.04 | 9.33 | — | — | 248.5 | — | — | — | — | — | 9.87 | 89.5 | 23.6 | |
| LlamaGen-dist#Params=111M, Epochs=502025.03 | 9.41 | — | — | 127.44 | — | — | — | — | — | — | — | — | |
| DenseModel Size=Medium, Model=DiT-XL/2, Compression Factor (CF)=2x2025.12 | 9.62 | — | — | 121.5 | — | — | — | — | — | 6.85 | — | — | |
| LlamaGen-sft#Params=111M, Epochs=502025.03 | 10.03 | — | — | 116.37 | — | — | — | — | — | — | — | — | |
| RiskyDiffTarget Model=DenseNet-1212025.12 | 10.2 | — | — | — | — | — | — | — | — | — | — | — | |
| RiskyDiffTarget Model=EfficientNet-B22025.12 | 10.4 | — | — | — | — | — | — | — | — | — | — | — | |
| BLASTModel Size=Medium, Model=DiT-XL/2, Compression Factor (CF)=2x2025.12 | 10.45 | — | — | 111.05 | — | — | — | — | — | 6.72 | — | — | |
| RiskyDiffTarget Model=ViT-B/162025.12 | 11.5 | — | — | — | — | — | — | — | — | — | — | — | |
| RiskyDiffTarget Model=ResNet-502025.12 | 12.4 | — | — | — | — | — | — | — | — | — | — | — | |
| Late-expansion VPTime-Conditioning=Late-expansion VP2026.04 | 12.46 | — | — | 294.8 | — | — | — | — | — | 13.41 | 92.1 | 18.6 | |
| Parallel DecodingNFE=642026.07 | 13.07 | — | — | — | — | — | — | — | — | — | — | — | |
| Original LDM2026.02 | 13.28 | — | — | — | — | — | — | — | — | — | — | — | |
| AdvDiffTarget Model=DenseNet-1212025.12 | 13.5 | — | — | — | — | — | — | — | — | — | — | — | |
| AdvDiffTarget Model=ViT-B/162025.12 | 13.8 | — | — | — | — | — | — | — | — | — | — | — | |
| FHSNFE=642026.07 | 14.31 | — | — | — | — | — | — | — | — | — | — | — | |
| CustomMark2026.02 | 14.73 | — | — | — | — | — | — | — | — | — | — | — | |
| AdvDiffTarget Model=EfficientNet-B22025.12 | 14.9 | — | — | — | — | — | — | — | — | — | — | — | |
| TokenTrace2026.02 | 14.98 | — | — | — | — | — | — | — | — | — | — | — | |
| IGAN2026.01 | 15.08 | — | — | 68.25 | — | — | — | 9.41 | 95.6 | — | — | — | |
| AdvDiffTarget Model=ResNet-502025.12 | 15.7 | — | — | — | — | — | — | — | — | — | — | — | |
| VQGAN#Params=1.4B, Epochs=3002025.03 | 15.78 | — | — | 74.3 | — | — | — | — | — | — | — | — | |
| t_orthoTime-Conditioning=t_ortho2026.04 | 17.52 | — | — | 305.18 | — | — | — | — | — | 18.48 | 93.3 | 11.3 | |
| ProMark2026.02 | 17.63 | — | — | — | — | — | — | — | — | — | — | — | |
| time embeddingTime-Conditioning=time embedding2026.04 | 18.07 | — | — | 308.31 | — | — | — | — | — | 20.45 | 93.2 | 11 | |
| VQGAN#Params=227M, Epochs=3002025.03 | 18.65 | — | — | 80.4 | — | — | — | — | — | — | — | — | |
| ContraGANBatch size=256, Source=This paper's implementation2020.06 | 19.443 | 7.7 | — | — | — | — | — | — | — | — | — | — | |
| BigGANBatch size=256, Source=This paper's implementation2020.06 | 21.072 | — | — | — | — | — | — | — | — | — | — | — | |
| BigGAN2026.01 | 22.77 | — | — | 38.05 | — | — | — | 8.49 | 91.4 | — | — | — | |
| AdvDiffuserTarget Model=ViT-B/162025.12 | 23.1 | — | — | — | — | — | — | — | — | — | — | — | |
| AdvDiffuserTarget Model=EfficientNet-B22025.12 | 25.4 | — | — | — | — | — | — | — | — | — | — | — | |
| AdvDiffuserTarget Model=DenseNet-1212025.12 | 26.4 | — | — | — | — | — | — | — | — | — | — | — | |
| SNGAN2026.01 | 26.79 | — | — | 32.25 | — | — | — | 9.38 | 91.3 | — | — | — | |
| AdvDiffuserTarget Model=ResNet-502025.12 | 27.1 | — | — | — | — | — | — | — | — | — | — | — | |
| SAGAN2026.01 | 34.73 | — | — | 29.85 | — | — | — | 9.47 | 85.5 | — | — | — |