Density Estimation on ImageNet 64x64 (test)
2.94Bits Per Sub-PixelstNCE-s+DSM
Evaluation results
| Method | Links | |
|---|---|---|
| stNCE-s+DSMAnti-aliasing=true, Augmentation=Horizontal flips, Discreteness=Continuous, Type=Estimate, Single NFE=true2026.05 | 2.94 | |
| TarFlowAnti-aliasing=false, Augmentation=Horizontal flips, Discreteness=Uniform, Type=Normalized, Single NFE=true2026.05 | 2.99 | |
| NFDMforward process parameterization=Gaussian2024.04 | 3.2 | |
| NFDMforward process parameterization=non-Gaussian2024.04 | 3.2 | |
| NFDMAnti-aliasing=false, Augmentation=Horizontal flips, Discreteness=Uniform*, Type=Normalized, Single NFE=false2026.05 | 3.2 | |
| Flow Matching2024.04 | 3.31 | |
| FMAnti-aliasing=true, Augmentation=None, Discreteness=Uniform*, Type=Normalized, Single NFE=false2026.05 | 3.31 | |
| DenseFlowTIME (GPU hours)=2242021.06 | 3.35 | |
| NDM2024.04 | 3.35 | |
| Dual SMAnti-aliasing=true, Augmentation=Horizontal flips, Discreteness=Continuous, Type=Estimate, Single NFE=true2026.05 | 3.36 | |
| VDMType=Diff, Evaluation Protocol=variational bound2021.07 | 3.4 | |
| Perceiver AR2023.05 | 3.4 | |
| MEGABYTE2023.05 | 3.4 | |
| VDM2024.04 | 3.4 | |
| VDMAnti-aliasing=false, Augmentation=None, Discreteness=Discrete, Type=Upper bound, Single NFE=false2026.05 | 3.4 | |
| DistAugPARAM=152.0M2021.06 | 3.42 | |
| Combiner2023.05 | 3.42 | |
| Routing TransformerType=AR2021.07 | 3.43 | |
| Routing Transformer2023.05 | 3.43 | |
| Sparse TransformerModel type=AR, Params=152M, Sampling=D2020.11 | 3.44 | |
| Sparse TransformerPARAM=152.0M, TIME (GPU hours)=7days2021.06 | 3.44 | |
| 2-resolution MRCNFPARAM=2.0M, TIME (GPU hours)=42.32021.06 | 3.44 | |
| Sparse TransformerType=AR2021.07 | 3.44 | |
| Improved DDPMType=Diff2021.07 | 3.51 | |
| PixelSNAILArchitecture Category=Autoregressive2019.02 | 3.52 | |
| SPNArchitecture Category=Autoregressive2019.02 | 3.52 | |
| SPNModel type=AR, Params=150M, Sampling=D2020.11 | 3.52 | |
| Very Deep VAEModel type=VAE, Params=125M, Depth=75, Sampling=12020.11 | 3.52 | |
| Very Deep VAEType=VAE2021.07 | 3.52 | |
| SPNType=AR2021.07 | 3.52 | |
| SPNPARAM=150.0M2021.06 | 3.53 | |
| Improved DDPM2024.04 | 3.54 | |
| I-DDPMAnti-aliasing=false, Augmentation=None, Discreteness=Continuous, Type=Upper bound, Single NFE=false2026.05 | 3.54 | |
| 3-resolution MRCNFPARAM=2.0M, TIME (GPU hours)=35.42021.06 | 3.55 | |
| PixelCNNResolution=64x642017.03 | 3.57 | |
| Gated PixelCNNArchitecture Category=Autoregressive2019.02 | 3.57 | |
| Gated PixelCNNModel type=AR, Params=177M*, Sampling=D2020.11 | 3.57 | |
| Gated PixelCNNTIME (GPU hours)=602021.06 | 3.57 | |
| PixelCNNType=AR2021.07 | 3.57 | |
| PixelCNNAnti-aliasing=false, Augmentation=None, Discreteness=Discrete, Type=Normalized, Single NFE=false2026.05 | 3.57 | |
| PixelRNNArchitecture Category=Autoregressive2019.02 | 3.63 | |
| PixelRNNResolution=64x642017.03 | 3.64 | |
| Flow++Architecture Category=Flow-based, Dequantization Strategy=Variational2019.02 | 3.69 | |
| MACOWArchitecture Category=Flow-based, Model Variation=+var, Dequantization Strategy=Variational2019.02 | 3.69 | |
| Flow++variational dequantization=true2019.06 | 3.69 | |
| Flow++Model type=Flow, Params=73M, Sampling=12020.11 | 3.69 | |
| MaCowPARAM=122.5M2021.06 | 3.69 | |
| Flow++PARAM=73.5M2021.06 | 3.69 | |
| Flow++Type=Flow2021.07 | 3.69 | |
| Multiscale PixelCNNResolution=64x642017.03 | 3.7 | |
| Parallel MultiscaleArchitecture Category=Autoregressive2019.02 | 3.7 | |
| MACOWArchitecture Category=Flow-based, Model Variation=+fine-grained2019.02 | 3.75 | |
| Residual Flow2019.06 | 3.757 | |
| MACOWArchitecture Category=Flow-based, Model Variation=Original (Org)2019.02 | 3.78 | |
| Wavelet FlowPARAM=96.0M, TIME (GPU hours)=8222021.06 | 3.78 | |
| GlowArchitecture Category=Flow-based2019.02 | 3.81 | |
| Glow2019.06 | 3.81 | |
| GlowModel type=Flow, Sampling=12020.11 | 3.81 | |
| GlowPARAM=111.1M2021.06 | 3.81 | |
| GlowType=Flow2021.07 | 3.81 | |
| GlowEpochs=6002025.12 | 3.81 | |
| EmergingEpochs=600, Convolution Type=Emerging convolutions2025.12 | 3.81 | |
| GlowAnti-aliasing=false, Augmentation=None, Discreteness=Continuous, Type=Normalized, Single NFE=true2026.05 | 3.81 | |
| RNODEGPUs=4, PARAM=2.0M, TIME (GPU hours)=256.42021.06 | 3.83 | |
| QuadEpochs=600, Convolution Type=3x3 invertible convolutions, Coupling Type=Quad-coupling2025.12 | 3.8514 | |
| 3 x 3Epochs=600, Convolution Type=3x3 invertible convolutions2025.12 | 3.8946 | |
| Real NVPResolution=64x642017.03 | 3.98 | |
| Real NVPArchitecture Category=Flow-based2019.02 | 3.98 | |
| Real NVP2019.06 | 3.98 | |
| RealNVPPARAM=96.0M2021.06 | 3.98 | |
| Conv. DRAWResolution=64x642017.03 | 4.1 |