Image Generation on ImageNet (val)
329.41Inception ScoreDense
Evaluation results
| Method | Links | |||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| DensePruning Strategy=Dense, Sparsity (%)=0%, MACs=2.37, Latency (ms)=83.60, Speedup=1.00×, Sampling steps=20, Backbone=DiT-XL/22026.03 | 329.41 | — | 4.49 | — | — | 87 | 48 | — | — | — | — | — | — | |
| FP BaselineModel Architecture=RAR-XXL, Bit Width=FP2026.01 | 328.87 | — | 1.51 | 5.13 | — | 81 | — | — | — | — | — | — | — | |
| PTQ4ARVGModel Architecture=RAR-XXL, Bit Width=W8A82026.01 | 321.03 | — | 1.61 | 5.28 | — | 82 | — | — | — | — | — | — | — | |
| VARKDTeacher -> Student=XL -> L, Base Student Architecture=ARPG-L2026.06 | 313 | — | 2.19 | — | — | 81 | 57 | — | — | — | — | — | — | |
| ARPG-L + SeqKDStudent Architecture=ARPG-L, Distillation=SeqKD2026.06 | 310.2 | — | 2.22 | — | — | 80 | 58 | — | — | — | — | — | — | |
| SeqKDTeacher -> Student=XXL -> L, Base Student Architecture=ARPG-L2026.06 | 310.2 | — | 2.22 | — | — | 80 | 58 | — | — | — | — | — | — | |
| GKDTeacher -> Student=XL -> L, Base Student Architecture=ARPG-L2026.06 | 310.1 | — | 2.26 | — | — | 81 | 58 | — | — | — | — | — | — | |
| FP BaselineModel Architecture=VAR-d20, Bit Width=FP2026.01 | 309.09 | — | 2.85 | 7.66 | — | 83 | — | — | — | — | — | — | — | |
| ARPG-LStudent Architecture=ARPG-L, Distillation=None2026.06 | 309 | — | 2.3 | — | — | 80 | 58 | — | — | — | — | — | — | |
| ARPG-L2026.06 | 309 | — | 2.3 | — | — | 80 | 58 | — | — | — | — | — | — | |
| ARPG-L + GKDStudent Architecture=ARPG-L, Distillation=GKD2026.06 | 307.8 | — | 2.19 | — | — | 81 | 59 | — | — | — | — | — | — | |
| GKDTeacher -> Student=XXL -> L, Base Student Architecture=ARPG-L2026.06 | 307.8 | — | 2.19 | — | — | 81 | 59 | — | — | — | — | — | — | |
| FP BaselineModel Architecture=VAR-d30, Bit Width=FP2026.01 | 307.24 | — | 2.03 | 8.72 | — | 81 | — | — | — | — | — | — | — | |
| SeqKDTeacher -> Student=XL -> L, Base Student Architecture=ARPG-L2026.06 | 305.2 | — | 2.28 | — | — | 80 | 59 | — | — | — | — | — | — | |
| Full Precision BaselineBit Width=FP, Model Scale=MAR-H2026.01 | 303.7 | — | 1.55 | — | — | 81 | — | — | — | — | — | — | — | |
| FP BaselineModel Architecture=RAR-L, Bit Width=FP2026.01 | 303 | — | 1.76 | 6.03 | — | 81 | — | — | — | — | — | — | — | |
| ARPG-L + VARKDStudent Architecture=ARPG-L, Distillation=VARKD2026.06 | 301.3 | — | 2.15 | — | — | 80 | 59 | — | — | — | — | — | — | |
| VARKDTeacher -> Student=XXL -> L, Base Student Architecture=ARPG-L2026.06 | 301.3 | — | 2.15 | — | — | 80 | 59 | — | — | — | — | — | — | |
| KDTeacher -> Student=XL -> L, Base Student Architecture=ARPG-L2026.06 | 296.1 | — | 2.2 | — | — | 80 | 59 | — | — | — | — | — | — | |
| MAR LargeDepth=32, Params=479 M, Epochs=4002024.12 | 296 | — | 1.78 | — | — | — | — | — | — | — | — | — | — | |
| Full Precision BaselineBit Width=FP, Model Scale=MAR-L2026.01 | 296 | — | 1.78 | — | — | 81 | — | — | — | — | — | — | — | |
| PTQ4ARVGBit Width=W8A8, Model Scale=MAR-H2026.01 | 294.66 | — | 1.67 | — | — | 79 | — | — | — | — | — | — | — | |
| ARPG-L + KDStudent Architecture=ARPG-L, Distillation=KD2026.06 | 293.1 | — | 2.21 | — | — | 80 | 58 | — | — | — | — | — | — | |
| KDTeacher -> Student=XXL -> L, Base Student Architecture=ARPG-L2026.06 | 293.1 | — | 2.21 | — | — | 80 | 58 | — | — | — | — | — | — | |
| PTQ4ARVGModel Architecture=RAR-L, Bit Width=W8A82026.01 | 291.55 | — | 1.9 | 6.34 | — | 81 | — | — | — | — | — | — | — | |
| OmniQuantModel Architecture=RAR-L, Bit Width=W8A82026.01 | 289.61 | — | 2.16 | 6.66 | — | 79 | — | — | — | — | — | — | — | |
| OmniQuantModel Architecture=RAR-XXL, Bit Width=W8A82026.01 | 288.1 | — | 2.35 | 6.43 | — | 77 | — | — | — | — | — | — | — | |
| MAR+CCFG=true, Generation Order=collapse order, Fine-tuning=false2025.11 | 284.78 | — | 5.956 | — | 2.321 | 82.6 | 56.6 | — | — | — | — | — | — | |
| SVDQuantModel Architecture=RAR-L, Bit Width=W8A82026.01 | 284.59 | — | 1.95 | 6.38 | — | 80 | — | — | — | — | — | — | — | |
| PTQ4ARVGBit Width=W8A8, Model Scale=MAR-L2026.01 | 284.02 | — | 1.92 | — | — | 80 | — | — | — | — | — | — | — | |
| TinyMAR-D16Depth=32->16, Params=277 M, Epochs=402024.12 | 283.4 | — | 2.28 | — | — | — | — | — | — | — | — | — | — | |
| MAR BaseDepth=24, Params=208 M, Epochs=4002024.12 | 281.7 | — | 2.31 | — | — | — | — | — | — | — | — | — | — | |
| Full Precision BaselineBit Width=FP, Model Scale=MAR-B2026.01 | 281.7 | — | 2.31 | — | — | 82 | — | — | — | — | — | — | — | |
| MARCFG=true, Generation Order=random, Fine-tuning=false2025.11 | 281.48 | — | 5.997 | — | 2.33 | 82.2 | 57.1 | — | — | — | — | — | — | |
| CMARCFG=true, Generation Order=collapse order, Fine-tuning=true2025.11 | 280.55 | — | 5.928 | — | 2.238 | 81.8 | 57.6 | — | — | — | — | — | — | |
| PTQ4ARVGBit Width=W8A8, Model Scale=MAR-B2026.01 | 279.97 | — | 2.36 | — | — | 82 | — | — | — | — | — | — | — | |
| SmoothQuantModel Architecture=RAR-XXL, Bit Width=W8A82026.01 | 278.89 | — | 2.09 | 5.91 | — | 77 | — | — | — | — | — | — | — | |
| SiT-XL/2Depth=28, Params=675 M, Epochs=1,4002024.12 | 277.5 | — | 2.06 | — | — | — | — | — | — | — | — | — | — | |
| PTQ4ARVGModel Architecture=VAR-d30, Bit Width=W8A82026.01 | 277.05 | — | 3.27 | 14.4 | — | 81 | — | — | — | — | — | — | — | |
| SVDQuantModel Architecture=RAR-XXL, Bit Width=W8A82026.01 | 276.52 | — | 2.14 | 5.46 | — | 75 | — | — | — | — | — | — | — | |
| DDPM/250Backbone=DiT-XL/2, Latency (s)=2.508, Speedup=1.00×2026.03 | 275.65 | — | 2.23 | 4.57 | — | 82 | 58 | — | — | — | — | — | — | |
| VA-VAEGenerative Model=Light DiT, Generative Paradigm=Non-autoregressive, Semantics Distillation=true, CFG Scale=Optimal2025.07 | 275.1 | — | 2.86 | — | — | — | — | — | — | — | — | — | — | |
| MAETokGenerative Model=Light DiT, Generative Paradigm=Non-autoregressive, Semantics Distillation=true, CFG Scale=Optimal2025.07 | 273.3 | — | 3.92 | — | — | — | — | — | — | — | — | — | — | |
| SODABackbone=DiT-XL/2, Sampling Mode=DDPM-based, Latency (s)=1.778, Speedup=1.41×2026.03 | 272.01 | — | 2.21 | 4.72 | — | 82 | 58 | — | — | — | — | — | — | |
| SODABackbone=DiT-XL/2, Sampling Mode=DDPM-based, Latency (s)=1.651, Speedup=1.52×2026.03 | 271.77 | — | 2.23 | 4.76 | — | 82 | 59 | — | — | — | — | — | — | |
| QuaRotModel Architecture=RAR-L, Bit Width=W8A82026.01 | 271.37 | — | 2.35 | 7.42 | — | 79 | — | — | — | — | — | — | — | |
| ∆-DiTBackbone=DiT-XL/2, Speedup=1.52×2026.03 | 271.03 | — | 2.31 | — | — | — | — | — | — | — | — | — | — | |
| OS+Model Architecture=VAR-d30, Bit Width=W8A82026.01 | 269.86 | — | 3.3 | 15.89 | — | 81 | — | — | — | — | — | — | — | |
| OmniQuantModel Architecture=VAR-d30, Bit Width=W8A82026.01 | 268.92 | — | 3.35 | 15.34 | — | 81 | — | — | — | — | — | — | — | |
| Diff-ESPruning Strategy=OBS, Sparsity (%)=31.25%, MACs=1.66, Latency (ms)=69.42, Speedup=1.20×, Sampling steps=20, Backbone=DiT-XL/22026.03 | 268.66 | — | 3.63 | — | — | 81 | 52 | — | — | — | — | — | — | |
| SVDQuantModel Architecture=VAR-d30, Bit Width=W8A82026.01 | 268.55 | — | 3.34 | 14.87 | — | 80 | — | — | — | — | — | — | — | |
| l-DeTokGenerative Model=MAR, Generative Paradigm=Autoregressive, Semantics Distillation=false, CFG Scale=Optimal2025.07 | 266.5 | — | 2.43 | — | — | — | — | — | — | — | — | — | — | |
| PTQ4ARVGModel Architecture=RAR-XXL, Bit Width=W6A62026.01 | 266.39 | — | 2.41 | 5.7 | — | 77 | — | — | — | — | — | — | — | |
| MAR-VAEGenerative Model=MAR, Generative Paradigm=Autoregressive, Semantics Distillation=false, CFG Scale=Optimal2025.07 | 265.3 | — | 3.71 | — | — | — | — | — | — | — | — | — | — | |
| PTQ4ARVGModel Architecture=VAR-d20, Bit Width=W8A82026.01 | 263.86 | — | 3.36 | 11.17 | — | 82 | — | — | — | — | — | — | — | |
| Full Precision BaselineBit Width=FP, Model Variant=PAR-3B-16x2026.01 | 262.5 | — | 2.88 | — | — | 82 | — | — | — | — | — | — | — | |
| RepQ*Model Architecture=VAR-d30, Bit Width=W8A82026.01 | 262.34 | — | 3.51 | 16.05 | — | 79 | — | — | — | — | — | — | — | |
| PTQ4ARVGBit Width=W6A6, Model Scale=MAR-H2026.01 | 261.35 | — | 2.62 | — | — | 75 | — | — | — | — | — | — | — | |
| OS+Model Architecture=VAR-d20, Bit Width=W8A82026.01 | 260.96 | — | 3.4 | 11.25 | — | 81 | — | — | — | — | — | — | — | |
| QuaRotModel Architecture=VAR-d20, Bit Width=W8A82026.01 | 260.52 | — | 4.02 | 11.2 | — | 81 | — | — | — | — | — | — | — | |
| SmoothQuantModel Architecture=VAR-d20, Bit Width=W8A82026.01 | 260.02 | — | 3.5 | 12.51 | — | 81 | — | — | — | — | — | — | — | |
| SD-VAEGenerative Model=MAR, Generative Paradigm=Autoregressive, Semantics Distillation=false, CFG Scale=Optimal2025.07 | 259.8 | — | 4.64 | — | — | — | — | — | — | — | — | — | — | |
| VAEtype=standard VAE2026.04 | 259.168 | — | — | — | — | — | 64.9 | 3.237 | — | 74.1 | — | — | — | |
| RepQ*Model Architecture=VAR-d20, Bit Width=W8A82026.01 | 258.9 | — | 3.46 | 11.98 | — | 79 | — | — | — | — | — | — | — | |
| LlamaGen-L + SeqKDStudent Architecture=LlamaGen-L, Distillation=SeqKD2026.06 | 258.1 | — | 3.1 | — | — | 83 | 52 | — | — | — | — | — | — | |
| SeqKDTeacher -> Student=XL -> L, Base Student Architecture=LLamaGen-L2026.06 | 258.1 | — | 3.1 | — | — | 83 | 52 | — | — | — | — | — | — | |
| l-DeTokGenerative Model=RasterAR, Generative Paradigm=Autoregressive, Semantics Distillation=false, CFG Scale=Optimal2025.07 | 257.7 | — | 4.46 | — | — | — | — | — | — | — | — | — | — | |
| Full Precision BaselineBit Width=FP, Model Variant=PAR-3B-4x2026.01 | 255.5 | — | 2.29 | — | — | 82 | — | — | — | — | — | — | — | |
| PTQ4ARVGBit Width=W6A6, Model Scale=MAR-L2026.01 | 254.13 | — | 3.12 | — | — | 76 | — | — | — | — | — | — | — | |
| l-DeTok + DistillationGenerative Model=MAR, Generative Paradigm=Autoregressive, Semantics Distillation=true, CFG Scale=Optimal2025.07 | 254.1 | — | 2.52 | — | — | — | — | — | — | — | — | — | — | |
| Diff-ESPruning Strategy=Wanda, Sparsity (%)=12.50%, MACs=2.13, Latency (ms)=78.87, Speedup=1.06×, Sampling steps=20, Backbone=DiT-XL/22026.03 | 253.79 | — | 5.28 | — | — | 79 | 50 | — | — | — | — | — | — | |
| RepQ*Model Architecture=RAR-L, Bit Width=W8A82026.01 | 253.16 | — | 2.55 | 7.34 | — | 76 | — | — | — | — | — | — | — | |
| OS+Model Architecture=RAR-L, Bit Width=W8A82026.01 | 252.98 | — | 2.7 | 7.94 | — | 67 | — | — | — | — | — | — | — | |
| SVDQuantModel Architecture=VAR-d20, Bit Width=W8A82026.01 | 252.17 | — | 3.8 | 11.87 | — | 81 | — | — | — | — | — | — | — | |
| MosaicDiffPruning Strategy=OBS, Sparsity (%)=31.46%, MACs=1.66, Latency (ms)=68.91, Speedup=1.22×, Sampling steps=20, Backbone=DiT-XL/22026.03 | 251.23 | — | 4.01 | — | — | 79 | 53 | — | — | — | — | — | — | |
| SD3-VAEGenerative Model=MAR, Generative Paradigm=Autoregressive, Semantics Distillation=false, CFG Scale=Optimal2025.07 | 250 | — | 6.46 | — | — | — | — | — | — | — | — | — | — | |
| PTQ4ARVGBit Width=W6A6, Model Scale=MAR-B2026.01 | 249.14 | — | 2.99 | — | — | 78 | — | — | — | — | — | — | — | |
| l-DeTokGenerative Model=RandomAR, Generative Paradigm=Autoregressive, Semantics Distillation=false, CFG Scale=Optimal2025.07 | 248.9 | — | 5.22 | — | — | — | — | — | — | — | — | — | — | |
| OS+Bit Width=W8A8, Model Scale=MAR-H2026.01 | 248.27 | — | 2.84 | — | — | 73 | — | — | — | — | — | — | — | |
| GKDTeacher -> Student=XXL -> L, Base Student Architecture=LLamaGen-L2026.06 | 248.1 | — | 2.88 | — | — | 82 | 55 | — | — | — | — | — | — | |
| LlamaGen-L + GKDStudent Architecture=LlamaGen-L, Distillation=GKD2026.06 | 248 | — | 2.93 | — | — | 82 | 55 | — | — | — | — | — | — | |
| GKDTeacher -> Student=XL -> L, Base Student Architecture=LLamaGen-L2026.06 | 248 | — | 2.93 | — | — | 82 | 55 | — | — | — | — | — | — | |
| SmoothQuantModel Architecture=VAR-d30, Bit Width=W8A82026.01 | 247.74 | — | 4.37 | 18.19 | — | 76 | — | — | — | — | — | — | — | |
| VARKDTeacher -> Student=XXL -> L, Base Student Architecture=LLamaGen-L2026.06 | 246.7 | — | 2.91 | — | — | 83 | 54 | — | — | — | — | — | — | |
| KDTeacher -> Student=XXL -> L, Base Student Architecture=LLamaGen-L2026.06 | 246.6 | — | 3.01 | — | — | 83 | 54 | — | — | — | — | — | — | |
| SmoothQuantModel Architecture=RAR-L, Bit Width=W8A82026.01 | 245.99 | — | 2.73 | 7.29 | — | 76 | — | — | — | — | — | — | — | |
| OS+Model Architecture=RAR-XXL, Bit Width=W8A82026.01 | 245.23 | — | 2.67 | 6.23 | — | 77 | — | — | — | — | — | — | — | |
| LlamaGen-L + KDStudent Architecture=LlamaGen-L, Distillation=KD2026.06 | 244.5 | — | 3.02 | — | — | 82 | 54 | — | — | — | — | — | — | |
| KDTeacher -> Student=XL -> L, Base Student Architecture=LLamaGen-L2026.06 | 244.5 | — | 3.02 | — | — | 82 | 54 | — | — | — | — | — | — | |
| MAETokGenerative Model=SiT, Generative Paradigm=Non-autoregressive, Semantics Distillation=true, CFG Scale=Optimal2025.07 | 243.2 | — | 4.77 | — | — | — | — | — | — | — | — | — | — | |
| l-DeTok + DistillationGenerative Model=Light DiT, Generative Paradigm=Non-autoregressive, Semantics Distillation=true, CFG Scale=Optimal2025.07 | 243 | — | 2.18 | — | — | — | — | — | — | — | — | — | — | |
| LlamaGen-L + VARKDStudent Architecture=LlamaGen-L, Distillation=VARKD2026.06 | 242.7 | — | 2.83 | — | — | 83 | 55 | — | — | — | — | — | — | |
| VARKDTeacher -> Student=XL -> L, Base Student Architecture=LLamaGen-L2026.06 | 242.7 | — | 2.83 | — | — | 83 | 55 | — | — | — | — | — | — | |
| RAE-ARBase Representation=SigLIP2, modification=+Norm+Noise2026.04 | 241.768 | — | — | — | — | — | 33.5 | 6.091 | — | 86.9 | — | — | — | |
| OmniQuantModel Architecture=VAR-d20, Bit Width=W8A82026.01 | 239.3 | — | 4.14 | 11.73 | — | 77 | — | — | — | — | — | — | — | |
| SeqKDTeacher -> Student=XXL -> L, Base Student Architecture=LLamaGen-L2026.06 | 238.6 | — | 3.15 | — | — | 83 | 52 | — | — | — | — | — | — | |
| RepQ*Model Architecture=RAR-XXL, Bit Width=W8A82026.01 | 233.35 | — | 3.25 | 6.45 | — | 77 | — | — | — | — | — | — | — | |
| SVDQuantBit Width=W8A8, Model Scale=MAR-H2026.01 | 233.13 | — | 4.21 | — | — | 71 | — | — | — | — | — | — | — | |
| Diff-ESPruning Strategy=OBS, Sparsity (%)=43.75%, MACs=1.42, Latency (ms)=61.20, Speedup=1.37×, Sampling steps=20, Backbone=DiT-XL/22026.03 | 232.73 | — | 4.25 | — | — | 77 | 57 | — | — | — | — | — | — | |
| l-DeTok + DistillationGenerative Model=SiT, Generative Paradigm=Non-autoregressive, Semantics Distillation=true, CFG Scale=Optimal2025.07 | 232.6 | — | 3.4 | — | — | — | — | — | — | — | — | — | — |