Image Generation on ImageNet 256x256 (train)
1.13FIDRAE+DiTDH-XL/2
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| RAE+DiTDH-XL/2NFE=50 × 2, Space=DINOv2, Params=839M, Epochs=8002026.06 | 1.13 | — | 262.6 | — | — | — | — | |
| RAEModel Space=Latent Space, NFE=50×2, Epochs=800, #Params=415 + 839M, GFLOPs=107 + 146, requires external models=true2025.10 | 1.28 | — | 262.9 | — | — | — | — | |
| SiT-XL/2 + REGEpochs=800, CFG=true, External=true2026.01 | 1.36 | 4.25 | 299.4 | 77 | 66 | — | — | |
| REGEpochs=800, CFG=True2026.03 | 1.36 | 4.25 | 299.4 | 77 | 66 | — | — | |
| SiD2 patch 1Representation Space=Pixel, Model Category=Diffusion Model, Maximum noise level (tmax)=0.72026.04 | 1.38 | — | — | — | — | — | — | |
| SiT-XL/2 + REPAEpochs=800, CFG=true, External=true2026.01 | 1.42 | 4.7 | 311.4 | 80 | 63 | — | — | |
| REPAModel Space=Latent Space, NFE=434, Epochs=800, #Params=84 + 675M, GFLOPs=312 + 119, requires external models=true2025.10 | 1.42 | — | 305.7 | 80 | 65 | — | — | |
| REPANFE=250×1.7, #Params=675M2025.03 | 1.42 | — | — | — | — | — | — | |
| SiT-XL/2 + VAE-REPAEpochs=800, CFG=true, External=false2026.01 | 1.52 | 4.63 | 316.2 | 82 | 62 | — | — | |
| DiverseDiT (Ours)Epochs=200, CFG=True2026.03 | 1.52 | 4.23 | 282.72 | 81 | 66 | — | — | |
| MAR-HNFE=256×2, #Params=943M2025.03 | 1.55 | — | — | — | — | — | — | |
| eMIGM-HNFE=128×1.4, #Params=942M2025.03 | 1.57 | — | — | — | — | — | — | |
| SiT-XL/2 + SRAEpochs=800, CFG=true, External=true2026.01 | 1.58 | 4.65 | 305.7 | 80 | 65 | — | — | |
| EPG-G/16Model Space=Pixel Space, NFE=75, Epochs=1600, #Params=1391M, GFLOPs=321, class-balanced sampling=true2025.10 | 1.58 | — | 298.4 | 80 | 63 | — | — | |
| MDTv2-XL/2Epochs=1080, CFG=True2026.03 | 1.58 | 4.52 | 314.7 | 79 | 65 | — | — | |
| SRAEpochs=800, CFG=True2026.03 | 1.58 | 4.65 | 311.4 | 80 | 63 | — | — | |
| FlowAR-Htype=scale-wise autoregressive, params=1.9B2024.12 | 1.65 | — | 296.5 | 83 | 60 | — | — | |
| E2E-REPAEpochs=800, CFG=True2026.03 | 1.69 | 4.17 | 219.3 | 77 | 67 | — | — | |
| EPG-G/16Model Space=Pixel Space, NFE=75, Epochs=1600, #Params=1391M, GFLOPs=3212025.10 | 1.7 | — | 297.8 | 80 | 63 | — | — | |
| SiD2Model Space=Pixel Space, GFLOPs=1372025.10 | 1.72 | — | — | — | — | — | — | |
| eMIGM-LNFE=128×1.4, #Params=478M2025.03 | 1.72 | — | — | — | — | — | — | |
| DiffiTCFG=true, External=false2026.01 | 1.73 | — | 276.5 | 80 | 62 | — | — | |
| DiffiTNFE=250×2, #Params=561M2025.03 | 1.73 | — | — | — | — | — | — | |
| EPG-G/16Model Space=Pixel Space, NFE=75, Epochs=600, #Params=1391M, GFLOPs=321, class-balanced sampling=true2025.10 | 1.75 | — | 275.1 | 80 | 62 | — | — | |
| MAR-LNFE=256×2, #Params=479M2025.03 | 1.78 | — | — | — | — | — | — | |
| MAR-LRepresentation Space=Latent, Model Category=Autoregressive Model, # Param.=479M+66M, Maximum noise level (tmax)=0.52026.04 | 1.78 | — | — | — | — | — | — | |
| REPAEpochs=800, CFG=True2026.03 | 1.8 | 4.5 | 284 | 81 | 61 | — | — | |
| EPG-XXL/16Model Space=Pixel Space, NFE=75, Epochs=600, #Params=789M, GFLOPs=176, class-balanced sampling=true2025.10 | 1.81 | — | 294.6 | 80 | 61 | — | — | |
| JiT-G/16Model Space=Pixel Space, NFE=191, Epochs=600, #Params=2B, GFLOPs=383, class-balanced sampling=true2025.10 | 1.82 | — | 292.6 | 79 | 62 | — | — | |
| SiT w. ViT-5-XLBackbone=ViT-5-XL, Training steps=7M, Classifier-free guidance=1.52026.02 | 1.84 | — | 282.73 | 83 | 60 | — | — | |
| SiT-XL/2 + VAE-REPAEpochs=400, CFG=true, External=false2026.01 | 1.84 | 4.5 | 306.5 | 82 | 61 | — | — | |
| JiT-H/16Model Space=Pixel Space, NFE=191, Epochs=600, #Params=953M, GFLOPs=182, class-balanced sampling=true2025.10 | 1.86 | — | 303.4 | 78 | 62 | — | — | |
| iMF-L/2NFE=1, Space=SD-VAE, Params=409M, Epochs=6402026.06 | 1.86 | — | — | — | — | — | — | |
| EPG-XXL/16Model Space=Pixel Space, NFE=75, Epochs=800, #Params=789M, GFLOPs=1762025.10 | 1.87 | — | 287.6 | 80 | 63 | — | — | |
| DiverseDiT (Ours)Epochs=80, CFG=True2026.03 | 1.89 | 4.44 | 276.85 | 81 | 66 | — | — | |
| FlowAR-Ltype=scale-wise autoregressive, params=589M2024.12 | 1.9 | — | 281.4 | 83 | 57 | — | — | |
| DeCo-XL/16NFE=100 × 2, Space=pixel, Params=682M, Epochs=3202026.06 | 1.9 | — | 303 | — | — | — | — | |
| VAR-d30NFE=10×2, #Params=2B2025.03 | 1.92 | — | — | — | — | — | — | |
| VAR-d30Family=VAR, #Params=2.0B, Epochs=200-3502026.05 | 1.92 | — | 323.1 | 82 | 59 | — | — | |
| PixNerd-XL/16Model Space=Pixel Space, NFE=100×2, Epochs=320, #Params=700M, GFLOPs=1342025.10 | 1.93 | — | 297 | 79 | 59 | — | — | |
| PixNerd-XLRepresentation Space=Pixel, Model Category=Diffusion Model, # Param.=700M, Maximum noise level (tmax)=0.72026.04 | 1.93 | — | — | — | — | — | — | |
| REPARepresentation Space=Latent, Model Category=Diffusion Model, # Param.=675M+86M, Maximum noise level (tmax)=0.52026.04 | 1.94 | — | — | — | — | — | — | |
| MIND-XL-GFamily=Diff.+Dis., #Params=~715M, Epochs=1600, GigaTok=true2026.05 | 1.95 | — | 293.78 | 75 | 67 | — | — | |
| REPAEpochs=200, CFG=True2026.03 | 1.96 | 4.49 | 264 | 82 | 60 | — | — | |
| VAR-d30type=scale-wise autoregressive, params=2.0B2024.12 | 1.97 | — | 323.1 | 82 | 59 | — | — | |
| DispLossEpochs=800, CFG=True2026.03 | 1.97 | — | — | — | — | — | — | |
| MIND-XLFamily=Diff.+Dis., #Params=~715M, Epochs=10002026.05 | 1.97 | — | 297.93 | 78 | 61 | — | — | |
| SiT-XL/2 + VAE-REPAEpochs=200, CFG=true, External=false2026.01 | 1.98 | 4.53 | 284.5 | 83 | 59 | — | — | |
| PixelFlowModel Space=Pixel Space, #Params=677M, GFLOPs=29092025.10 | 1.98 | — | 282.1 | 81 | 60 | — | — | |
| TF-H/16+FD-lossNFE=4×1, Space=DINOv2, Params=1.1B, Epochs=802026.06 | 1.98 | — | 261.6 | — | — | — | — | |
| DiT-XL/2 + iSSDEpochs=1400+0.8, CFG=true, External=true2026.01 | 2.02 | 4.22 | 250 | 81 | 60 | — | — | |
| eMIGM-HNFE=16×1.2, #Params=942M2025.03 | 2.02 | — | — | — | — | — | — | |
| TF-L/16+FD-lossNFE=4×1, Space=DINOv2, Params=515M, Epochs=802026.06 | 2.02 | — | 259.3 | — | — | — | — | |
| EPG-XL/16Model Space=Pixel Space, NFE=75, Epochs=800, #Params=583M, GFLOPs=1282025.10 | 2.04 | — | 283.2 | 80 | 61 | — | — | |
| IBQ-XXLFamily=AR, #Params=2.1B, Epochs=4502026.05 | 2.05 | — | 286.73 | 83 | 57 | — | — | |
| SiTtype=flow matching, params=675M2024.12 | 2.06 | — | 270.3 | 82 | 59 | — | — | |
| SiT w. ViT-XLBackbone=ViT-XL, Training steps=7M, Classifier-free guidance=1.52026.02 | 2.06 | — | 277.5 | 83 | 59 | — | — | |
| SiT-XL/2Epochs=1400, CFG=true, External=false2026.01 | 2.06 | 4.5 | 270.3 | 82 | 59 | — | — | |
| SiT-XL/2Model Space=Latent Space, NFE=250×2, Epochs=1400, #Params=84 + 675M, GFLOPs=312 + 1192025.10 | 2.06 | — | 277.5 | 83 | 59 | — | — | |
| SiT-XLNFE=250×2, #Params=675M2025.03 | 2.06 | — | — | — | — | — | — | |
| SiT-XL/2Epochs=1400, CFG=True2026.03 | 2.06 | 4.5 | 270.3 | 82 | 59 | — | — | |
| SiT-XLRepresentation Space=Latent, Model Category=Diffusion Model, # Param.=675M+86M, Maximum noise level (tmax)=0.52026.04 | 2.06 | — | — | — | — | — | — | |
| MIND-B-GFamily=Diff.+Dis., #Params=~130M, Epochs=1600, GigaTok=true2026.05 | 2.06 | — | 268.03 | 78 | 62 | — | — | |
| SiT-XL/2 (cfg=1.50)NFE=250 × 2, Space=SD-VAE, Params=675M, Epochs=14002026.06 | 2.06 | — | 270.3 | — | — | — | — | |
| VAR-d24NFE=10×2, #Params=1B2025.03 | 2.09 | — | — | — | — | — | — | |
| VAR-d24Family=VAR, #Params=1.0B, Epochs=200-3502026.05 | 2.09 | — | 312.9 | 82 | 59 | — | — | |
| Large-DiTNFE=250×2, #Params=3B2025.03 | 2.1 | — | — | — | — | — | — | |
| VDM++Model Space=Pixel Space, NFE=250×2, #Params=2.46B, GFLOPs=5552025.10 | 2.12 | — | 278.1 | — | — | — | — | |
| IBQ-XLFamily=AR, #Params=1.1B, Epochs=4002026.05 | 2.14 | — | 278.99 | 83 | 56 | — | — | |
| LlamaGen-3Btype=token-wise autoregressive, params=3.1B2024.12 | 2.18 | — | 263.3 | 81 | 58 | — | — | |
| LlamaGen-3BFamily=AR, #Params=3.1B, Epochs=3002026.05 | 2.18 | — | 263.33 | 81 | 58 | — | — | |
| MIND-BFamily=Diff.+Dis., #Params=~130M, Epochs=10002026.05 | 2.18 | — | 258.46 | 80 | 58 | — | — | |
| MIND-BFamily=Diff.+Dis., #Params=~130M, Epochs=8962026.05 | 2.21 | — | 260.54 | 80 | 58 | — | — | |
| eMIGM-LNFE=16×1.2, #Params=478M2025.03 | 2.22 | — | — | — | — | — | — | |
| DiT-XL/2-GModel=Diff., Label fraction=100%, #Params=675M2023.02 | 2.27 | 4.6 | 278.24 | 0.83 | 0.57 | — | — | |
| DiTtype=diffusion, params=675M2024.12 | 2.27 | — | 278.2 | 83 | 57 | — | — | |
| DiT w. ViT-XLBackbone=ViT-XL, Training steps=7M, Classifier-free guidance=1.52026.02 | 2.27 | — | 278.24 | 83 | 57 | — | — | |
| DiT-XL/2Epochs=1400, CFG=true, External=false2026.01 | 2.27 | 4.6 | 278.2 | 83 | 57 | — | — | |
| DiT-XL/2Model Space=Latent Space, NFE=250×2, Epochs=1400, #Params=84 + 675M, GFLOPs=312 + 1192025.10 | 2.27 | — | 278.2 | 83 | 57 | — | — | |
| DiT-XL/2NFE=250×2, #Params=675M2025.03 | 2.27 | — | — | — | — | — | — | |
| DiT-XL/2Epochs=1400, CFG=True2026.03 | 2.27 | 4.6 | 278.2 | 83 | 57 | — | — | |
| DiT-XLRepresentation Space=Latent, Model Category=Diffusion Model, # Param.=675M+86M, Maximum noise level (tmax)=0.52026.04 | 2.27 | — | — | — | — | — | — | |
| DiT-XL/2 (cfg=1.50)NFE=250 × 2, Space=SD-VAE, Params=675M, Epochs=14002026.06 | 2.27 | — | 278.2 | — | — | — | — | |
| iMF-M/2NFE=1, Space=SD-VAE, Params=174M, Epochs=6402026.06 | 2.27 | — | — | — | — | — | — | |
| MaskDiTEpochs=1600, CFG=true, External=true2026.01 | 2.28 | 5.67 | 276.6 | 80 | 61 | — | — | |
| MaskDiTModel Space=Latent Space, NFE=79×2, Epochs=1600, #Params=84 + 675M, GFLOPs=312 + 1192025.10 | 2.28 | — | 276.6 | 80 | 61 | — | — | |
| Large-DiTNFE=250×2, #Params=7B2025.03 | 2.28 | — | — | — | — | — | — | |
| DIFFUSSM-XL-GNFE=250×2, #Params=660M2025.03 | 2.28 | — | — | — | — | — | — | |
| MakDiTEpochs=1600, CFG=True2026.03 | 2.28 | 5.67 | 276.7 | 80 | 62 | — | — | |
| U-ViT-H/2Objective=SD, Training iterations=500K, Convolution block=without conv, Weight mode=ft-EMA, p_uncond=0.1, #Params=501M + 84M (AE), GFLOPS=133 + 312 (AE)2022.09 | 2.29 | 5.68 | 263.88 | 82 | 57 | — | — | |
| Supervised baselineModel=Diff., Backbone=U-ViT-Huge, Label fraction=100%, #Params=585M2023.02 | 2.29 | 5.68 | 263.88 | 0.82 | 0.57 | 1.75 | — | |
| U-ViT-H/2Epochs=240, CFG=true, External=false2026.01 | 2.29 | 5.68 | 263.9 | 82 | 57 | — | — | |
| U-ViT-H/2NFE=50×2, #Params=501M2025.03 | 2.29 | — | — | — | — | — | — | |
| StyleGAN-XLModel=GAN, Label fraction=100%2023.02 | 2.3 | 4.02 | 265.12 | 0.78 | 0.53 | — | — | |
| StyleGANtype=GAN, params=166M2024.12 | 2.3 | — | 265.1 | 78 | 53 | — | — | |
| StyleGAN-XLNFE=1×22025.03 | 2.3 | — | — | — | — | — | — | |
| MAR-BNFE=256×2, #Params=208M2025.03 | 2.31 | — | — | — | — | — | — | |
| eMIGM-BNFE=128×1.35, #Params=208M2025.03 | 2.32 | — | — | — | — | — | — | |
| iTARFlow-BRepresentation Space=Latent, Model Category=Normalizing Flow, # Param.=770M+86M, Maximum noise level (tmax)=0.52026.04 | 2.32 | — | — | — | — | — | — | |
| USP DiT-XL/2Model Space=Latent Space, NFE=250×2, Epochs=240, #Params=84 + 675M, GFLOPs=312 + 1192025.10 | 2.33 | — | 267 | — | — | — | — |