Image Generation on ImageNet 1K (val)
1.34FIDRecTok
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| RecTokParams=176M, GFlops=44.42025.12 | 1.34 | — | — | — | — | — | |
| RAEParams=395M, GFlops=128.92025.12 | 1.51 | — | — | — | — | — | |
| DeTokParams=176M, GFlops=44.42025.12 | 1.86 | — | — | — | — | — | |
| AdamWBase Optimizer=AdamW, Backbone=SiT-XL/2, Rank=512, Optimizer Mem. (GB)=5.1, Model Mem. (GB)=2.5, Training Time=20.8 h, Resolution=256x256, Iterations=400K, Inference Steps=2502024.11 | 1.9 | — | — | — | — | — | |
| AdafactorBase Optimizer=Adafactor, Backbone=SiT-XL/2, Rank=512, Optimizer Mem. (GB)=2.5, Model Mem. (GB)=2.5, Training Time=25.5 h, Resolution=256x256, Iterations=400K, Inference Steps=2502024.11 | 1.9 | — | — | — | — | — | |
| COAPBase Optimizer=Adafactor, Backbone=SiT-XL/2, Rank=512, Optimizer Mem. (GB)=1.5 (-40%), Model Mem. (GB)=2.5, Training Time=+7%, Resolution=256x256, Iterations=400K, Inference Steps=2502024.11 | 2.1 | — | — | — | — | — | |
| VARModel=VAR, Configuration=d242026.03 | 2.11 | 311 | — | 0.82 | 59 | 1 | |
| VA-VAEParams=70M, GFlops=3102025.12 | 2.17 | — | — | — | — | — | |
| COAPBase Optimizer=AdamW, Backbone=SiT-XL/2, Rank=512, Optimizer Mem. (GB)=2.6 (-49%), Model Mem. (GB)=2.5, Training Time=+14%, Resolution=256x256, Iterations=400K, Inference Steps=2502024.11 | 2.2 | — | — | — | — | — | |
| MAETokParams=176M, GFlops=54.22025.12 | 2.21 | — | — | — | — | — | |
| CoDeModel=VAR2026.03 | 2.26 | 297 | — | 0.8 | 60 | 2.9 | |
| DDIMModel=DiT, Steps=502026.03 | 2.27 | 241 | — | — | — | 1 | |
| GaLoreBase Optimizer=AdamW, Backbone=SiT-XL/2, Rank=512, Optimizer Mem. (GB)=2.6 (-49%), Model Mem. (GB)=2.5, Training Time=+38%, Resolution=256x256, Iterations=400K, Inference Steps=2502024.11 | 2.3 | — | — | — | — | — | |
| CIARModel=DiT2026.03 | 2.32 | 240 | — | — | — | 4.9 | |
| CIARModel=VAR2026.03 | 2.34 | 290 | — | 0.81 | 60 | 2.6 | |
| FP (Full Precision)Bit Width=FP, Model Scale=PAR-XXL-4x2026.01 | 2.35 | 263.2 | — | 0.82 | — | — | |
| FP (Full Precision)Bit Width=FP, Model Scale=PAR-XL-4x2026.01 | 2.61 | 259.2 | — | 0.82 | — | — | |
| VARModel=VAR, Configuration=d202026.03 | 2.61 | 301 | — | 0.83 | 56 | 1.65 | |
| SVDQuantBit Width=W8A8, Model Scale=PAR-XXL-4x2026.01 | 2.91 | 222.74 | 6.83 | 0.78 | — | — | |
| PTQ4ARVGBit Width=W8A8, Model Scale=PAR-XXL-4x2026.01 | 2.95 | 232.16 | 6.6 | 0.79 | — | — | |
| GaLoreBase Optimizer=Adafactor, Backbone=SiT-XL/2, Rank=512, Optimizer Mem. (GB)=1.5 (-40%), Model Mem. (GB)=2.5, Training Time=+33%, Resolution=256x256, Iterations=400K, Inference Steps=2502024.11 | 3 | — | — | — | — | — | |
| OmniQuantBit Width=W8A8, Model Scale=PAR-XXL-4x2026.01 | 3.05 | 224.87 | 6.73 | 0.78 | — | — | |
| DDIMModel=DiT, Steps=252026.03 | 3.18 | 232 | — | — | — | 2 | |
| OmniQuantBit Width=W8A8, Model Scale=PAR-XL-4x2026.01 | 3.55 | 215.71 | 7.78 | 0.78 | — | — | |
| PTQ4ARVGBit Width=W8A8, Model Scale=PAR-XL-4x2026.01 | 3.55 | 219.5 | 7.71 | 0.79 | — | — | |
| SVDQuantBit Width=W8A8, Model Scale=PAR-XL-4x2026.01 | 3.8 | 213.98 | 7.79 | 0.78 | — | — | |
| DDIMModel=DiT, Steps=202026.03 | 3.81 | 221 | — | — | — | 2.5 | |
| ArcVQ-VAECodebook size (K)=1024, Generative Framework=LDM2026.05 | 6.79 | 172.3 | — | 0.75 | 52 | — | |
| CVQ-VAECodebook size (K)=1024, Generative Framework=LDM2026.05 | 6.87 | — | — | — | — | — | |
| MoVQCodebook size (K)=1024, Generative Framework=MaskGIT2026.05 | 7.13 | 138.3 | — | 0.75 | 57 | — | |
| RQVAECodebook size (K)=16384, Generative Framework=RQ-Transformer2026.05 | 7.55 | 134 | — | — | — | — | |
| DDIMModel=DiT, Steps=122026.03 | 7.8 | 184 | — | — | — | 4.17 | |
| SD-VAEParams=84M, GFlops=4452025.12 | 8.3 | — | — | — | — | — | |
| VQGAN-LCCodebook size (K)=100000, Generative Framework=LDM2026.05 | 8.36 | 191.5 | — | 0.71 | 49 | — | |
| VQGAN-EMACodebook size (K)=16384, Generative Framework=LDM2026.05 | 9.13 | — | — | — | — | — | |
| VQGAN-FCCodebook size (K)=16384, Generative Framework=LDM2026.05 | 9.78 | — | — | — | — | — | |
| PTQ4ARVGBit Width=W6A6, Model Scale=PAR-XXL-4x2026.01 | 11.05 | 119.19 | 7.17 | 0.63 | — | — | |
| PTQ4ARVGBit Width=W6A6, Model Scale=PAR-XL-4x2026.01 | 12.87 | 113.33 | 6.8 | 0.62 | — | — | |
| SQ+STWQBit Width=W6A6, Model Scale=PAR-XL-4x2026.01 | 13.26 | 107.33 | 7.09 | 0.59 | — | — | |
| SQ+STWQBit Width=W6A6, Model Scale=PAR-XXL-4x2026.01 | 14.69 | 94.27 | 7.73 | 0.59 | — | — | |
| COAPRank Ratio=2, Optimizer Mem. (GB)=1.8, Training Time=+13%, Base Optimizer=AdamW2024.11 | 16.2 | — | — | — | — | — | |
| GaLoreRank Ratio=2, Optimizer Mem. (GB)=2.0, Training Time=+21%, Base Optimizer=AdamW2024.11 | 17.8 | — | — | — | — | — | |
| AdamWOptimizer Mem. (GB)=3.0, Training Time=310.3 h2024.11 | 18 | — | — | — | — | — | |
| COAPRank Ratio=2.2, Optimizer Mem. (GB)=1.3, Training Time=+7%, Base Optimizer=Adafactor2024.11 | 18.3 | — | — | — | — | — | |
| SVDQuantBit Width=W6A6, Model Scale=PAR-XXL-4x2026.01 | 18.8 | 101.71 | 20.56 | 0.63 | — | — | |
| Derfbackbone=DiT-XL/22025.12 | 18.92 | — | — | — | — | — | |
| Derfmodel=DiT-XL/22025.12 | 18.92 | — | — | — | — | — | |
| SVDQuantBit Width=W6A6, Model Scale=PAR-XL-4x2026.01 | 19.52 | 102.03 | 18.19 | 0.6 | — | — | |
| LNbackbone=DiT-XL/22025.12 | 19.94 | — | — | — | — | — | |
| LNmodel=DiT-XL/22025.12 | 19.94 | — | — | — | — | — | |
| RepQ*+STWQBit Width=W6A6, Model Scale=PAR-XXL-4x2026.01 | 20.74 | 82.35 | 12.96 | 0.54 | — | — | |
| RMSNormmodel=DiT-XL/22025.12 | 20.76 | — | — | — | — | — | |
| DyTbackbone=DiT-XL/22025.12 | 20.83 | — | — | — | — | — | |
| DyTmodel=DiT-XL/22025.12 | 20.83 | — | — | — | — | — | |
| RepQ*+STWQBit Width=W6A6, Model Scale=PAR-XL-4x2026.01 | 23.04 | 72.96 | 9.93 | 0.5 | — | — | |
| GaLoreRank Ratio=2.2, Optimizer Mem. (GB)=1.8, Training Time=+18%, Base Optimizer=Adafactor2024.11 | 23.3 | — | — | — | — | — | |
| XDLMsteps=16, conditioning=standard2026.02 | 25.774 | 43.903 | — | — | — | — | |
| UDLMsteps=16, conditioning=standard2026.02 | 26.242 | 41.801 | — | — | — | — | |
| MDLMsteps=16, conditioning=standard2026.02 | 28.785 | 44.656 | — | — | — | — | |
| UDLMsteps=8, conditioning=standard2026.02 | 30.144 | 38.832 | — | — | — | — | |
| XDLMsteps=8, conditioning=standard2026.02 | 34.109 | 36.964 | — | — | — | — | |
| GIDDsteps=16, conditioning=standard2026.02 | 35.403 | 35.698 | — | — | — | — | |
| AdafactorOptimizer Mem. (GB)=2.2, Training Time=336.9 h2024.11 | 38.7 | — | — | — | — | — | |
| Derfbackbone=DiT-L/42025.12 | 43.94 | — | — | — | — | — | |
| Derfmodel=DiT-L/42025.12 | 43.94 | — | — | — | — | — | |
| RMSNormmodel=DiT-L/42025.12 | 45.02 | — | — | — | — | — | |
| DyTbackbone=DiT-L/42025.12 | 45.66 | — | — | — | — | — | |
| DyTmodel=DiT-L/42025.12 | 45.66 | — | — | — | — | — | |
| LNbackbone=DiT-L/42025.12 | 45.91 | — | — | — | — | — | |
| LNmodel=DiT-L/42025.12 | 45.91 | — | — | — | — | — | |
| MDLMsteps=8, conditioning=standard2026.02 | 47.732 | 29.178 | — | — | — | — | |
| UDLMsteps=4, conditioning=standard2026.02 | 49.861 | 27.049 | — | — | — | — | |
| NaLaSiT2025.06 | 53.08 | 27.63 | 8.94 | 0.43 | 62 | — | |
| EfficientSiT2025.06 | 53.57 | 27.26 | 9.01 | 0.43 | 61 | — | |
| XDLMsteps=4, conditioning=standard2026.02 | 54.085 | 24.829 | — | — | — | — | |
| GIDDsteps=8, conditioning=standard2026.02 | 54.933 | 24.297 | — | — | — | — | |
| SiT2025.06 | 58.61 | 24.31 | 9.25 | 0.41 | 59 | — | |
| NaLaDiT2025.06 | 61.64 | 23.24 | 12.4 | 0.4 | 58 | — | |
| DiG2025.06 | 62.06 | 22.81 | 11.77 | 0.39 | 56 | — | |
| Derfbackbone=DiT-B/42025.12 | 63.23 | — | — | — | — | — | |
| Derfmodel=DiT-B/42025.12 | 63.23 | — | — | — | — | — | |
| DyTbackbone=DiT-B/42025.12 | 63.94 | — | — | — | — | — | |
| DyTmodel=DiT-B/42025.12 | 63.94 | — | — | — | — | — | |
| LNbackbone=DiT-B/42025.12 | 64.93 | — | — | — | — | — | |
| LNmodel=DiT-B/42025.12 | 64.93 | — | — | — | — | — | |
| RMSNormmodel=DiT-B/42025.12 | 65.08 | — | — | — | — | — | |
| DiT2025.06 | 68.4 | — | — | — | — | — | |
| MDLMsteps=4, conditioning=standard2026.02 | 80.752 | 16.287 | — | — | — | — | |
| GIDDsteps=4, conditioning=standard2026.02 | 86.842 | 14.559 | — | — | — | — | |
| OmniQuantBit Width=W6A6, Model Scale=PAR-XL-4x2026.01 | 107.29 | 15.72 | 53.01 | 0.2 | — | — | |
| OmniQuantBit Width=W6A6, Model Scale=PAR-XXL-4x2026.01 | 113.72 | 15.15 | 73.68 | 0.18 | — | — | |
| FloraBase Optimizer=Adafactor, Backbone=SiT-XL/2, Rank=512, Optimizer Mem. (GB)=1.6 (-36%), Model Mem. (GB)=2.5, Training Time=+7%, Resolution=256x256, Iterations=400K, Inference Steps=2502024.11 | 115.2 | — | — | — | — | — | |
| OS+Bit Width=W8A8, Model Scale=PAR-XL-4x2026.01 | 128.72 | 7.31 | 93.63 | 0.07 | — | — | |
| SmoothQuantBit Width=W8A8, Model Scale=PAR-XL-4x2026.01 | 132.17 | 8.28 | 79.5 | 0.06 | — | — | |
| OS+Bit Width=W6A6, Model Scale=PAR-XL-4x2026.01 | 138.78 | 7.27 | 86.01 | 0.06 | — | — | |
| RepQ*Bit Width=W8A8, Model Scale=PAR-XL-4x2026.01 | 138.97 | 6.44 | 114.35 | 0.06 | — | — | |
| SmoothQuantBit Width=W6A6, Model Scale=PAR-XL-4x2026.01 | 146.52 | 8.08 | 69.36 | 0.05 | — | — | |
| ReLoRABase Optimizer=AdamW, Backbone=SiT-XL/2, Rank=512, Optimizer Mem. (GB)=3.6 (-29%), Model Mem. (GB)=3.7 (+48%), Training Time=+33%, Resolution=256x256, Iterations=400K, Inference Steps=2502024.11 | 151.8 | — | — | — | — | — | |
| LoRABase Optimizer=AdamW, Backbone=SiT-XL/2, Rank=512, Optimizer Mem. (GB)=3.6 (-29%), Model Mem. (GB)=3.7 (+48%), Training Time=+33%, Resolution=256x256, Iterations=400K, Inference Steps=2502024.11 | 151.9 | — | — | — | — | — | |
| RepQ*Bit Width=W6A6, Model Scale=PAR-XL-4x2026.01 | 157.07 | 5.82 | 132.66 | 0.05 | — | — |