Image Generation on ImageNet 256x256
478.6ISProMoE-XL-Flow
Evaluation Results
| Method | Links | |||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| ProMoE-XL-FlowTraining Steps=2M, # Activated Params.=675M, # Total Params.=1.568B, cfg=4.02025.10 | 478.6 | — | — | — | — | — | — | — | — | 11.95 | — | — | — | — | — | |
| DiffMoE-XL-FlowTraining Steps=2M, # Activated Params.=675M, # Total Params.=1.735B, cfg=4.02025.10 | 475.13 | — | — | — | — | — | — | — | — | 14.25 | — | — | — | — | — | |
| Dense-DiT-XL-FlowTraining Steps=2M, # Activated Params.=675M, # Total Params.=675M, cfg=4.02025.10 | 474.39 | — | — | — | — | — | — | — | — | 17.05 | — | — | — | — | — | |
| ProMoE-L-FlowTraining Steps=2M, # Activated Params.=458M, # Total Params.=1.063B, cfg=4.02025.10 | 466.96 | — | — | — | — | — | — | — | — | 14.05 | — | — | — | — | — | |
| DiffMoE-L-FlowTraining Steps=2M, # Activated Params.=458M, # Total Params.=1.163B, cfg=4.02025.10 | 465.6 | — | — | — | — | — | — | — | — | 15.7 | — | — | — | — | — | |
| Dense-DiT-L-FlowTraining Steps=2M, # Activated Params.=458M, # Total Params.=458M, cfg=4.02025.10 | 464.25 | — | — | — | — | — | — | — | — | 17.71 | — | — | — | — | — | |
| VARTokenizer=VQVAE, Training Epochs=350, #params=2.0B, Classifier-Free Guidance=true2025.04 | 365.4 | 1.8 | — | 83 | 57 | — | — | — | — | — | — | — | — | — | — | |
| VAREpochs=350, Classifier-Free Guidance (CFG)=true, Category=Autoregressive Models2025.12 | 365.4 | — | — | 83 | 57 | 1.8 | — | — | — | — | — | — | — | — | — | |
| VAREpochs=350, with CFG=true2026.06 | 365.4 | — | — | 83 | 57 | 1.8 | — | — | — | — | — | — | — | — | — | |
| MaskGITClass=Masked Modeling2023.12 | 355.6 | — | — | 83 | 44 | 4.02 | — | — | — | — | — | — | — | — | — | |
| Predictor–CorrectorBackbone=SoFlow-XL/2-cond, K=20, t_noise=0.52026.05 | 349.84 | — | 5.33 | 87 | 50 | 4.67 | — | — | — | — | — | — | — | — | — | |
| FRACTALMAR-HPARAMS=848M, STEPS=962026.02 | 348.9 | 6.15 | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| FractalMAR-HPre-training Tokenizer=-, Pre-training SSL Encoder=-, #Params=848M, #Epochs=6002026.03 | 348.9 | — | — | — | — | 6.15 | — | — | — | — | — | — | — | — | — | |
| FractalMAR-HParams=848M, Epochs=8002026.05 | 348.9 | — | — | 81 | 46 | 6.15 | — | — | — | — | — | — | — | — | — | |
| FractalMAR-HParams=848M, Epochs=600, Diffusion Space=Pixel-space2026.06 | 348.9 | — | — | 81 | 46 | 6.15 | — | — | — | — | — | — | — | — | — | |
| RQTransGenerator Type=AR, Generator #Params=1.4B, Classifier Guidance=True2026.05 | 337.5 | 3.89 | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| AliTok-LGenerator Type=AR, Generator #Params=318M, Classifier Guidance=True2026.05 | 326.2 | 1.38 | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| RQ-TransformerClass=Autoregressive2023.12 | 323.7 | — | — | — | — | 3.8 | — | — | — | — | — | — | — | — | — | |
| RQ-Tran.Type=AR, Params=3.8B2026.03 | 323.7 | 3.8 | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| VAREpochs=350, #Params=2B, Generation Resolution=256x256, Guidance Setting=w/o guidance2026.01 | 323.1 | 1.92 | — | 82 | 59 | — | — | — | — | — | — | — | — | — | — | |
| RandAR-XXLType=AR, Params=1.4B2026.03 | 322 | 2.15 | — | 79 | 62 | — | — | — | — | — | — | — | — | — | — | |
| REGEpochs=80, Classifier-Free Guidance (CFG)=true, Backbone=SiT-XL/2, Category=Leveraging VFM's Representations2025.12 | 321.4 | — | 4.49 | 76 | 63 | 1.86 | — | — | — | — | — | — | — | — | — | |
| REGEpochs=80, with CFG=true2026.06 | 321.4 | — | 4.49 | 76 | 63 | 1.86 | — | — | — | — | — | — | — | — | — | |
| SDE-DPM-Solver++ (2M)Schedule=Time-uniform, NFE=202026.01 | 320.83 | — | 7 | — | — | 7.65 | — | — | — | — | — | — | — | — | — | |
| REGLUEEpochs=160, Classifier-Free Guidance (CFG)=true, Backbone=SiT-XL/2, Category=Leveraging VFM's Representations2025.12 | 320.4 | — | 4.26 | 78 | 65 | 1.53 | — | — | — | — | — | — | — | — | — | |
| SDE-DPM-Solver++ (2M)Schedule=LAS, NFE=202026.01 | 320.18 | — | 6.18 | — | — | 6.67 | — | — | — | — | — | — | — | — | — | |
| MagViT-v2Tokenizer=LFQ tokenizers, Training Epochs=1080, #params=307M, Classifier-Free Guidance=true2025.04 | 319.4 | 1.78 | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| MagViTv2Epochs=1080, Classifier-Free Guidance (CFG)=true, Category=Autoregressive Models2025.12 | 319.4 | — | — | 83 | 57 | 1.78 | — | — | — | — | — | — | — | — | — | |
| MAGVIT-v2Generator Type=Mask, Generator #Params=307M, Classifier Guidance=True2026.05 | 319.4 | 1.78 | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| MagViTv2Epochs=1080, with CFG=true2026.06 | 319.4 | — | — | 83 | 57 | 1.78 | — | — | — | — | — | — | — | — | — | |
| Checkerboard-LScale Ratio=2x, Classifier-free guidance (cfg)=1.5, Params=343M, #Steps=17, Time (s)=0.522026.02 | 318.2 | — | — | 82 | 57 | 2.83 | — | — | — | — | — | — | — | — | — | |
| RandAR-XLType=AR, Params=775M2026.03 | 317.8 | 2.25 | — | 80 | 60 | — | — | — | — | — | — | — | — | — | — | |
| MDTV2Epochs=1080, with CFG=true2026.06 | 317.7 | — | 4.52 | 79 | 65 | 1.58 | — | — | — | — | — | — | — | — | — | |
| SiT-XL/2 + AHPACategory=Latent diffusion, Transformer, Epochs=800, CFG=true, Standalone (w/o Ext.)=true2026.05 | 317.5 | — | 4.12 | 81 | 64 | 1.4 | — | — | — | — | — | — | — | — | — | |
| SiT-XL/2 + SRA 2Category=Latent diffusion, Transformer, Epochs=800, CFG=true, Standalone (w/o Ext.)=true2026.05 | 316.2 | — | 4.63 | 82 | 62 | 1.52 | — | — | — | — | — | — | — | — | — | |
| SiT-XL/2 + oursTraining Steps=4M, Classifier-free Guidance (CFG)=Enabled2026.07 | 315.3 | — | 4.6 | 81 | 63 | 1.44 | — | — | — | — | — | — | — | — | — | |
| MDTv2Tokenizer=SD-VAE, Training Epochs=1080, #params=675M, Classifier-Free Guidance=true2025.04 | 314.7 | 1.58 | 4.52 | 79 | 65 | — | — | — | — | — | — | — | — | — | — | |
| MDTv2-XL/2Category=Transformer + U-Net hybrid, Epochs=1080, CFG=True2026.01 | 314.7 | — | 4.52 | 79 | 65 | 1.58 | — | — | — | — | — | — | — | — | — | |
| REGLUEEpochs=80, Classifier-Free Guidance (CFG)=true, Backbone=SiT-XL/2, Category=Leveraging VFM's Representations2025.12 | 313.1 | — | 4.25 | 78 | 64 | 1.61 | — | — | — | — | — | — | — | — | — | |
| VAR-d24Type=AR, Params=1.0B2026.03 | 312.9 | 2.09 | — | 82 | 59 | — | — | — | — | — | — | — | — | — | — | |
| MaskAlignEpochs=800, with CFG=true2026.06 | 312.9 | — | 4.31 | 78 | 67 | 1.35 | — | — | — | — | — | — | — | — | — | |
| Checkerboard-LScale Ratio=4x, Classifier-free guidance (cfg)=1.7, Params=343M, #Steps=17, Time (s)=0.522026.02 | 311.5 | — | — | 80 | 57 | 2.79 | — | — | — | — | — | — | — | — | — | |
| SiT-XL/2 + SRACategory=Training Acceleration, Epochs=800, CFG=True2026.01 | 311.4 | — | 4.65 | 80 | 63 | 1.58 | — | — | — | — | — | — | — | — | — | |
| SiT-XL/2 + REPACategory=Latent diffusion, Transformer, Epochs=800, CFG=true, Standalone (w/o Ext.)=false2026.05 | 311.4 | — | 4.7 | 80 | 63 | 1.42 | — | — | — | — | — | — | — | — | — | |
| SiT-XL/2 + SRATraining Steps=4M, Classifier-free Guidance (CFG)=Enabled2026.07 | 311.4 | — | 4.65 | 80 | 63 | 1.58 | — | — | — | — | — | — | — | — | — | |
| Self-Transcendence (SiT-XL/2)Category=Training Acceleration, Backbone=SiT-XL/2, Epochs=400, CFG=True2026.01 | 311.3 | — | 4.85 | 79 | 66 | 1.44 | — | — | — | — | — | — | — | — | — | |
| DDT-XL/2Pre-training Tokenizer=SD-VAE, Pre-training SSL Encoder=DINOv2, #Params=675+49M, #Epochs=4002026.03 | 310.6 | — | — | — | — | 1.26 | — | — | — | — | — | — | — | — | — | |
| DDT-XL/2Framework Category=Two-stage Frameworks with Aux Supervision (DINOv2), Aux. Token=SD-VAE, Params=675M+49M2026.03 | 310.6 | — | — | — | — | 1.26 | — | — | — | — | — | — | — | — | — | |
| DDT-XL/2Params=675+49M, Epochs=4002026.05 | 310.6 | — | — | 78 | 67 | 1.26 | — | — | — | — | — | — | — | — | — | |
| DDT-XLParams=675M, Epochs=400, GFLOPs=119, Diffusion Space=Latent-space2026.06 | 310.6 | — | — | 79 | 65 | 1.26 | — | — | — | — | — | — | — | — | — | |
| SemTok-AR-XXLType=AR, Params=1.2B2026.03 | 310.5 | 2.34 | — | 77 | 63 | — | — | — | — | — | — | — | — | — | — | |
| MaskAlignEpochs=80, with CFG=true2026.06 | 310 | — | 4.48 | 81 | 63 | 1.82 | — | — | — | — | — | — | — | — | — | |
| UNITE-XLFramework Category=Single-stage Frameworks, Aux. Token=Joint, Params=806M2026.03 | 309.9 | — | — | — | — | 1.75 | — | — | — | — | — | — | — | — | — | |
| SDE-DPM-Solver++ (2M)Schedule=LogSNR, NFE=202026.01 | 308.36 | — | 8.56 | — | — | 7.15 | — | — | — | — | — | — | — | — | — | |
| SiT-XL/2 + AHPACategory=Latent diffusion, Transformer, Epochs=400, CFG=true, Standalone (w/o Ext.)=true2026.05 | 308.2 | — | 4.15 | 81 | 63 | 1.56 | — | — | — | — | — | — | — | — | — | |
| PixelU-H/16Params=1168M, Epochs=320, NFE=100×2, GFLOPs=137, Diffusion Space=Pixel-space2026.06 | 306.86 | — | 5.29 | 78 | 63 | 1.77 | — | — | — | — | — | — | — | — | — | |
| HyperDiT-HParams=952M, Epochs=600, NFE=100x22026.05 | 306.5 | — | 4.73 | 80 | 64 | 1.56 | — | — | — | — | — | — | — | — | — | |
| dMALABackbone=SoFlow-XL/2-cond, K=20, σ=0.32026.05 | 306.25 | — | 5.81 | 80 | 62 | 2.17 | — | — | — | — | — | — | — | — | — | |
| PixelU-H/16Params=1168M, Epochs=600, NFE=100×2, GFLOPs=137, Diffusion Space=Pixel-space2026.06 | 305.88 | — | 5.04 | 79 | 64 | 1.63 | — | — | — | — | — | — | — | — | — | |
| REPATokenizer=SD-VAE, Training Epochs=800, #params=675M, Classifier-Free Guidance=true2025.04 | 305.7 | 1.28 | 4.68 | 79 | 64 | — | — | — | — | — | — | — | — | — | — | |
| SiT-XL/2 + REPANFE=250×2, space=latent, params=675M (49M), Gflops=119 (310)2026.01 | 305.7 | — | — | — | — | 1.42 | — | — | — | — | — | — | — | — | — | |
| SiT-XL/2 + REPACategory=Training Acceleration, Epochs=800, CFG=True2026.01 | 305.7 | — | 4.7 | 80 | 65 | 1.42 | — | — | — | — | — | — | — | — | — | |
| REPAEpochs=800, Classifier-Free Guidance (CFG)=true, Backbone=SiT-XL/2, Category=Leveraging VFM's Representations2025.12 | 305.7 | — | 4.7 | 80 | 65 | 1.42 | — | — | — | — | — | — | — | — | — | |
| REPA-XLTokenizer=SD-VAE, Training Epochs=800, Steps=250, #params=675M, Classifier-Free Guidance (CFG)=true2025.10 | 305.7 | 1.42 | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| REPA, SiT-XL/2Pre-training Tokenizer=SD-VAE, Pre-training SSL Encoder=DINOv2, #Params=675+49M, #Epochs=8002026.03 | 305.7 | — | — | — | — | 1.42 | — | — | — | — | — | — | — | — | — | |
| REPA-SiT-XL/2Framework Category=Two-stage Frameworks with Aux Supervision (DINOv2), Aux. Token=SD-VAE, Params=675M+49M2026.03 | 305.7 | — | — | — | — | 1.42 | — | — | — | — | — | — | — | — | — | |
| REPA-XL/2Generator Type=Diff., Generator #Params=675M, Classifier Guidance=True2026.05 | 305.7 | 1.42 | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SiT-XL/2 + SRACategory=Latent diffusion, Transformer, Epochs=800, CFG=true, Standalone (w/o Ext.)=true2026.05 | 305.7 | — | 4.65 | 80 | 63 | 1.58 | — | — | — | — | — | — | — | — | — | |
| REPA-XL/2Params=675+49M, Epochs=800, NFE=250x22026.05 | 305.7 | — | 4.7 | 80 | 64 | 1.42 | — | — | — | — | — | — | — | — | — | |
| SiT-XL/2+REPA# Params.=675M, NFE=250 × 22026.05 | 305.7 | — | — | — | — | 1.42 | — | — | — | — | — | — | — | — | — | |
| REPAEpochs=800, with CFG=true2026.06 | 305.7 | — | 4.7 | 80 | 65 | 1.42 | — | — | — | — | — | — | — | — | — | |
| REPA-XL/2Params=675M + 86M, Epochs=800, NFE=250×2, GFLOPs=119, Diffusion Space=Latent-space2026.06 | 305.7 | — | 4.7 | 80 | 64 | 1.42 | — | — | — | — | — | — | — | — | — | |
| SiT-XL/2 + REPATraining Steps=4M, Classifier-free Guidance (CFG)=Enabled2026.07 | 305.7 | — | 4.7 | 80 | 65 | 1.42 | — | — | — | — | — | — | — | — | — | |
| SemTok-AR-XLType=AR, Params=746M2026.03 | 305.6 | 2.54 | — | 78 | 60 | — | — | — | — | — | — | — | — | — | — | |
| SpectualAR-d20Generator Type=AR, Generator #Params=600M, Classifier Guidance=True2026.05 | 305.4 | 2.49 | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| SiT-XL/2 + Self-FlowTraining Steps=4M, Classifier-free Guidance (CFG)=Enabled2026.07 | 305.4 | — | 4.54 | 80 | 61 | 1.47 | — | — | — | — | — | — | — | — | — | |
| REGEpochs=160, Classifier-Free Guidance (CFG)=true, Backbone=SiT-XL/2, Category=Leveraging VFM's Representations2025.12 | 304.6 | — | 4.36 | 77 | 65 | 1.59 | — | — | — | — | — | — | — | — | — | |
| REGEpochs=160, with CFG=true2026.06 | 304.6 | — | 4.36 | 77 | 65 | 1.59 | — | — | — | — | — | — | — | — | — | |
| ProMoE-XL-FlowTraining Steps=2M, # Activated Params.=675M, # Total Params.=1.568B, cfg=1.52025.10 | 304.26 | — | — | — | — | — | — | — | — | 2.08 | — | — | — | — | — | |
| HyperDiT-XLParams=676M, Epochs=600, NFE=100x22026.05 | 304.2 | — | 4.76 | 80 | 62 | 1.63 | — | — | — | — | — | — | — | — | — | |
| MaskAlignEpochs=160, with CFG=true2026.06 | 304.1 | — | 4.37 | 79 | 65 | 1.56 | — | — | — | — | — | — | — | — | — | |
| DeCo-XL/16Pre-training Tokenizer=-, Pre-training SSL Encoder=DINOv2, #Params=682M, #Epochs=6002026.03 | 304 | — | — | — | — | 1.69 | — | — | — | — | — | — | — | — | — | |
| DeCo-XL/16Params=682M, Epochs=600, NFE=100x22026.05 | 304 | — | 4.59 | 79 | 63 | 1.69 | — | — | — | — | — | — | — | — | — | |
| DeCo-XL/16Params=682M, Epochs=600, NFE=100×2, Diffusion Space=Pixel-space2026.06 | 304 | — | 4.59 | 79 | 63 | 1.69 | — | — | — | — | — | — | — | — | — | |
| Self-Transcendence (LightningDiT-XL/1)Category=Training Acceleration, Backbone=LightningDiT-XL/1, Epochs=400, CFG=True2026.01 | 303.9 | — | 4.11 | 78 | 66 | 1.25 | — | — | — | — | — | — | — | — | — | |
| MARTokenizer=LDM, Training Epochs=800, #params=945M, Classifier-Free Guidance=true2025.04 | 303.7 | 1.55 | — | 81 | 62 | — | — | — | — | — | — | — | — | — | — | |
| MAREpochs=800, Classifier-Free Guidance (CFG)=true, Category=Autoregressive Models2025.12 | 303.7 | — | — | 81 | 62 | 1.55 | — | — | — | — | — | — | — | — | — | |
| MAREpochs=800, with CFG=true2026.06 | 303.7 | — | — | 81 | 62 | 1.55 | — | — | — | — | — | — | — | — | — | |
| JIT-HPARAMS=953M, STEPS=1002026.02 | 303.4 | 1.86 | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| JiT-H/16Pre-training Tokenizer=-, Pre-training SSL Encoder=-, #Params=953M, #Epochs=6002026.03 | 303.4 | — | — | — | — | 1.86 | — | — | — | — | — | — | — | — | — | |
| JiT-H/16Framework Category=Single-stage Frameworks, Aux. Token=-, Params=953M2026.03 | 303.4 | — | — | — | — | 1.86 | — | — | — | — | — | — | — | — | — | |
| REPATokenizer=E2E-VAE (Ours), Training Epochs=800, #params=675M, Classifier-Free Guidance=true2025.04 | 302.9 | 1.12 | 4.09 | 79 | 66 | — | — | — | — | — | — | — | — | — | — | |
| VAR-d20Type=Parallelized AR, Params.=600M2026.01 | 302.6 | — | — | 83 | 56 | 2.95 | — | — | — | — | — | — | — | — | — | |
| VAR-d20Type=AR, Params=600M2026.03 | 302.6 | 2.57 | — | 83 | 56 | — | — | — | — | — | — | — | — | — | — | |
| VAR-d20Generator Type=VAR, Generator #Params=310M, Classifier Guidance=True2026.05 | 302.6 | 2.57 | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| Checkerboard-LScale Ratio=2x, Classifier-free guidance (cfg)=1.4, Params=343M, #Steps=17, Time (s)=0.522026.02 | 302.5 | — | — | 81 | 56 | 2.72 | — | — | — | — | — | — | — | — | — | |
| SPHERE-LPARAMS=950M, STEPS=42026.02 | 301.8 | 4.76 | — | — | — | — | — | — | — | — | — | — | — | — | — | |
| dMALABackbone=SoFlow-XL/2-cond, K=10, σ=0.32026.05 | 300.04 | — | 5.05 | 79 | 65 | 1.7 | — | — | — | — | — | — | — | — | — | |
| PixNerd-XL/16Params=700M, Epochs=320, NFE=100x22026.05 | 300 | — | 4.54 | 80 | 60 | 1.95 | — | — | — | — | — | — | — | — | — | |
| PixNerd-XL/16Params=700M, Epochs=320, NFE=100×2, GFLOPs=134, Diffusion Space=Pixel-space2026.06 | 300 | — | 4.54 | 80 | 60 | 1.95 | — | — | — | — | — | — | — | — | — |