Class-conditional Image Generation on ImageNet 256x256 (test)
1.07FIDLightningDiT+IG
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| LightningDiT+IGEpochs=680, #Params=678M, CFG=true, Sampling Protocol=Balanced sampling2025.12 | 1.07 | 274.1 | 79 | 66 | — | — | |
| RAE (DiTDH)Epochs=800, #Params=839M, CFG=true, Sampling Protocol=Balanced sampling2025.12 | 1.13 | 262.6 | 78 | 67 | — | — | |
| RAE (DiTDH-XL/2)Type=Flow, Tokenizer=C, Loss=MSE, #Param=839M2026.04 | 1.13 | 262.6 | — | — | — | — | |
| REPA-EEpochs=800, #Params=675M, CFG=true, Sampling Protocol=Balanced sampling2025.12 | 1.15 | 304 | 79 | 66 | — | — | |
| xAREpochs=800, #Params=1.1B, CFG=true, Sampling Protocol=Balanced sampling2025.12 | 1.24 | 301.6 | 83 | 64 | — | — | |
| LightningDiT+IGEpochs=680, #Params=678M, CFG=false, Sampling Protocol=Balanced sampling2025.12 | 1.24 | 229.6 | 78 | 66 | — | — | |
| BitDance-H-1xType=Hybrid, Tokenizer=D, Loss=MSE, #Param=1B2026.04 | 1.24 | 304.4 | — | — | — | — | |
| DE-CMNFE=50, #Params=675M, Time=0.582026.02 | 1.26 | — | — | — | — | — | |
| DDTEpochs=400, #Params=675M, CFG=true, Sampling Protocol=Balanced sampling2025.12 | 1.26 | 310.6 | 79 | 65 | — | — | |
| LightningDiTEpochs=800, #Params=675M, CFG=true, Sampling Protocol=Balanced sampling2025.12 | 1.28 | 300.5 | 80 | 65 | — | — | |
| REPAEpochs=800, #Params=675M, CFG=true, Sampling Protocol=Balanced sampling2025.12 | 1.29 | 306.3 | 79 | 64 | — | — | |
| OURS, FLOP HEAVYPARAMS=399M, NFE=2562026.03 | 1.32 | 294.2 | 74 | 64 | 3.71 | — | |
| DE-CMNFE=2, #Params=675M, Time=0.0242026.02 | 1.33 | — | — | — | — | — | |
| LightningDiT-XL/1NFE=250 × 2, #Params=675M, Time=> 2.82, CFG weight=6.72026.02 | 1.35 | — | — | — | — | — | |
| SiD2, FLOP HEAVYPARAMS=397M, NFE=5122026.03 | 1.37 | 286.3 | 73 | 65 | 3.83 | — | |
| REPA-XL/2NFE=250 × 2, #Params=675M, Time=> 2.89, CFG weight=1.82026.02 | 1.42 | — | — | — | — | — | |
| OURS, SMALLPARAMS=399M, NFE=2562026.03 | 1.42 | 297 | 73 | 65 | 3.82 | — | |
| REPA (SiT-XL/2)Type=Flow, Tokenizer=C, Loss=MSE, #Param=675M2026.04 | 1.42 | 305.7 | — | — | — | — | |
| SiT-XL/2 + REPA, CFG=1.5Train Steps=4M, Params (M)=675, GFlops/Iter=140.5, Stage 2 PFlops=143.9K, Total PFlops=143.9K2025.10 | 1.42 | — | — | — | — | — | |
| RAE (DiTDH)Epochs=800, #Params=839M, CFG=false, Sampling Protocol=Balanced sampling2025.12 | 1.51 | 242.9 | 79 | 63 | — | — | |
| MARNFE=100, #Params=945M2026.02 | 1.55 | — | — | — | — | — | |
| MAREpochs=800, #Params=943M, CFG=true, Sampling Protocol=Balanced sampling2025.12 | 1.55 | 303.7 | 81 | 62 | — | — | |
| MAR-HLatency/s=9.97, Speedup=1.002025.12 | 1.55 | 303.7 | — | — | — | — | |
| TransDiff-HLatency/s=6.72, Speedup=1.002025.12 | 1.55 | 297.9 | — | — | — | — | |
| MARType=Hybrid, Tokenizer=C, Loss=MSE, #Param=943M2026.04 | 1.55 | 303.7 | — | — | — | — | |
| FACMNFE=4, #Params=675M, Time=0.0472026.02 | 1.56 | — | — | — | — | — | |
| TransDiff-H + Fast-ARDiffLatency/s=1.38, Speedup=4.882025.12 | 1.6 | 295 | — | — | — | — | |
| MAR-H + Fast-ARDiffLatency/s=2.23, Speedup=4.472025.12 | 1.61 | 304.3 | — | — | — | — | |
| TransDiff-LLatency/s=3.17, Speedup=1.002025.12 | 1.61 | 295.1 | — | — | — | — | |
| PIXELDIT-XLPARAMS=797M, NFE=1002026.03 | 1.61 | 292.7 | 78 | 64 | 4.68 | — | |
| TransDiff-L + Fast-ARDiffLatency/s=0.68, Speedup=4.692025.12 | 1.66 | 294.1 | — | — | — | — | |
| TransDiff-H + SDLatency/s=3.52, Speedup=1.912025.12 | 1.68 | 291.1 | — | — | — | — | |
| SiD2, SMALLPARAMS=397M, NFE=5122026.03 | 1.68 | 288.2 | 72 | 65 | 4.04 | — | |
| MAR-H + LazyMARLatency/s=4.24, Speedup=2.352025.12 | 1.69 | 299.2 | — | — | — | — | |
| DiT-XL/2 with CADSGuidance scale (w_CFG)=22023.10 | 1.7 | 268.77 | 77 | 64 | — | — | |
| DE-CMNFE=1, #Params=675M, Time=0.0112026.02 | 1.7 | — | — | — | — | — | |
| REPA-EEpochs=800, #Params=675M, CFG=false, Sampling Protocol=Balanced sampling2025.12 | 1.7 | 217.3 | 77 | 66 | — | — | |
| TransDiff-H + DMDLatency/s=3.48, Speedup=1.932025.12 | 1.71 | 289.9 | — | — | — | — | |
| VAREpochs=350, #Params=2.0B, CFG=true, Sampling Protocol=Balanced sampling2025.12 | 1.73 | 350.2 | 82 | 60 | — | — | |
| MAR-H + DMDLatency/s=4.17, Speedup=2.392025.12 | 1.73 | 301 | — | — | — | — | |
| MeanFlow-†NFE=2, #Params=675M, Time=0.048, distillation_reproduction=true2026.02 | 1.74 | — | — | — | — | — | |
| FACMNFE=1, #Params=675M, Time=0.0112026.02 | 1.76 | — | — | — | — | — | |
| MAR-L + Fast-ARDiffLatency/s=1.21, Speedup=4.382025.12 | 1.77 | 297.1 | — | — | — | — | |
| MAR-LLatency/s=5.31, Speedup=1.002025.12 | 1.78 | 296 | — | — | — | — | |
| MDT-G2023.10 | 1.79 | 283.01 | 81 | 61 | — | — | |
| TransDiff-L + DMDLatency/s=1.61, Speedup=1.972025.12 | 1.79 | 288.3 | — | — | — | — | |
| DiT-XL/2 with CADSGuidance scale (w_CFG)=2.252023.10 | 1.81 | 288.1 | 78 | 64 | — | — | |
| MAR-L + DMDLatency/s=1.99, Speedup=2.672025.12 | 1.81 | 295.5 | — | — | — | — | |
| GRN-GType=AR, Tokenizer=D, Loss=CE, #Param=2B2026.04 | 1.81 | 299 | — | — | — | — | |
| JIT-GPARAMS=2B2026.03 | 1.82 | 292.6 | 79 | 62 | — | — | |
| JiT-G/16Type=Flow, Tokenizer=C, Loss=MSE, #Param=2B2026.04 | 1.82 | 292.6 | — | — | — | — | |
| DiT-G++2023.10 | 1.83 | 281.53 | 78 | 64 | — | — | |
| JiT-H/16Type=Flow, Tokenizer=C, Loss=MSE, #Param=953M2026.04 | 1.86 | 303.4 | — | — | — | — | |
| TransDiff-L + SDLatency/s=1.75, Speedup=1.812025.12 | 1.88 | 283.6 | — | — | — | — | |
| RepTok-L + CFG=1.5Stage 1 PFlops=30.4K, Train Steps=460K, Params (M)=516, GFlops/Iter=25.0, Stage 2 PFlops=11.7K, Total PFlops=42.1K2025.10 | 1.88 | — | — | — | — | — | |
| VAR-d30NFE=10, #Params=2B, Time=12026.02 | 1.92 | — | — | — | — | — | |
| VAREpochs=350, #Params=2.0B, CFG=false, Sampling Protocol=Balanced sampling2025.12 | 1.92 | 323.1 | 82 | 59 | — | — | |
| VAR-d30Type=AR, Tokenizer=D, Loss=CE, #Param=2B2026.04 | 1.92 | 323.1 | — | — | — | — | |
| DiT-XL/2 with CADSGuidance scale (w_CFG)=2.52023.10 | 1.93 | 297.96 | 78 | 63 | — | — | |
| MAR-L + LazyMARLatency/s=2.29, Speedup=2.322025.12 | 1.93 | 297.4 | — | — | — | — | |
| sCMs-†NFE=2, #Params=675M, Time=0.024, distillation_reproduction=true2026.02 | 1.94 | — | — | — | — | — | |
| PIXELFLOW-XLPARAMS=677M, NFE=ADAPT.2026.03 | 1.98 | 282.1 | 81 | 60 | 5.83 | — | |
| SiT-XL/2Type=Flow, Tokenizer=C, Loss=MSE, #Param=675M2026.04 | 2.06 | 277.5 | — | — | — | — | |
| GRN-HType=AR, Tokenizer=D, Loss=CE, #Param=952M2026.04 | 2.06 | 316.1 | — | — | — | — | |
| SiT-XL/2 + CFG=1.5Train Steps=7M, Params (M)=675, GFlops/Iter=118.6, Stage 2 PFlops=212.5K, Total PFlops=212.5K2025.10 | 2.06 | — | — | — | — | — | |
| RepTok-LStage 1 PFlops=30.4K, Train Steps=460K, Params (M)=516, GFlops/Iter=25.0, Stage 2 PFlops=11.7K, Total PFlops=42.1K2025.10 | 2.06 | — | — | — | — | — | |
| VAR-d24Type=AR, Tokenizer=D, Loss=CE, #Param=1B2026.04 | 2.09 | 312.9 | — | — | — | — | |
| L-DiTNFE=250, #Params=3B, Time=4.12026.02 | 2.1 | — | — | — | — | — | |
| LightningDiTEpochs=800, #Params=675M, CFG=false, Sampling Protocol=Balanced sampling2025.12 | 2.11 | 207 | 77 | 65 | — | — | |
| VDM++2023.10 | 2.12 | 267.7 | — | — | — | — | |
| SiT-XL/2NFE=250 × 2, #Params=675M, Time=> 2.89, CFG weight=1.52026.02 | 2.15 | — | — | — | — | — | |
| PIXNERD-XL/16PARAMS=700M, NFE=1002026.03 | 2.15 | 297 | 79 | 59 | 4.55 | — | |
| RandAR-XXLType=AR, Tokenizer=D, Loss=CE, #Param=1.4B2026.04 | 2.15 | 322 | — | — | — | — | |
| LightningDiT-XL/1NFE=250, #Params=675M, Time=2.822026.02 | 2.17 | — | — | — | — | — | |
| LlamaGen-3BType=AR, Parameters=3.1B, cfg=1.652024.06 | 2.18 | 263.33 | 81 | 58 | — | — | |
| MeanFlowNFE=2, #Params=675M, Time=0.0242026.02 | 2.2 | — | — | — | — | — | |
| VA-VAEEpochs=800, Steps=2502026.02 | 2.23 | — | — | — | — | 74.3 | |
| DiT-XL/2-GGuidance scale (w_CFG)=1.52023.10 | 2.27 | 278.24 | 83 | 57 | — | — | |
| DiT-XL/2Type=Diffusion, Parameters=675M2024.06 | 2.27 | 278.2 | 83 | 57 | — | — | |
| DiT-XL/2NFE=250 × 2, #Params=675M, Time=> 3.75, CFG weight=1.52026.02 | 2.27 | — | — | — | — | — | |
| DiT-XL/2Type=Diff, Params=675M, Steps=2502025.12 | 2.27 | 278.2 | 83 | 57 | — | — | |
| DiT-XL/2Type=Diffusion, Tokenizer=C, Loss=MSE, #Param=675M2026.04 | 2.27 | 278.2 | — | — | — | — | |
| StyleGAN-XL2023.10 | 2.3 | 265.12 | 78 | 53 | — | — | |
| StyleGan-XLType=GAN, Parameters=166M2024.06 | 2.3 | 265.1 | 78 | 53 | — | — | |
| StyleGAN-XLNFE=1, #Params=166M, Time=0.32026.02 | 2.3 | — | — | — | — | — | |
| StyleGAN-XLType=GAN, Params=166M, Steps=12025.12 | 2.3 | 265.1 | 78 | 53 | — | — | |
| LlamaGen-3BType=AR, Parameters=3.1B, cfg=1.752024.06 | 2.32 | 280.1 | 82 | 56 | — | — | |
| LlamaGen-XXLType=AR, Parameters=1.4B, cfg=1.752024.06 | 2.34 | 253.9 | 80 | 59 | — | — | |
| LlamaGen-XXLType=AR, Tokenizer=D, Loss=CE, #Param=1.4B2026.04 | 2.34 | 253.9 | — | — | — | — | |
| MAREpochs=800, #Params=943M, CFG=false, Sampling Protocol=Balanced sampling2025.12 | 2.35 | 227.8 | 79 | 62 | — | — | |
| JiT-L/16Type=Flow, Tokenizer=C, Loss=MSE, #Param=459M2026.04 | 2.36 | 298.5 | — | — | — | — | |
| LV-RAEEpochs=800, Steps=2502026.02 | 2.42 | — | — | — | — | 58.2 | |
| SimpleDiffusion2023.10 | 2.44 | 256.3 | — | — | — | — | |
| VAR-d20Type=AR, Tokenizer=D, Loss=CE, #Param=600M2026.04 | 2.57 | 302.6 | — | — | — | — | |
| LlamaGen-XLType=AR, Parameters=775M, cfg=1.752024.06 | 2.62 | 244.08 | 80 | 57 | — | — | |
| LlamaGen-XLType=AR, Tokenizer=D, Loss=CE, #Param=775M2026.04 | 2.62 | 244.1 | — | — | — | — | |
| GRN-LType=AR, Tokenizer=D, Loss=CE, #Param=458M2026.04 | 2.64 | 314.8 | — | — | — | — | |
| NAR-XLType=Parallelized AR, Params=816M, Steps=312025.12 | 2.7 | 277.5 | 81 | 58 | — | — | |
| TFM-XL/2# Params.=676M, NFE=22026.03 | 2.77 | — | — | — | — | — | |
| MeanFlow-†NFE=1, #Params=675M, Time=0.041, distillation_reproduction=true2026.02 | 2.79 | — | — | — | — | — |