Density Estimation on CIFAR-10 (test)
2.48Bits/dimNFDM
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| NFDMforward process parameterization=non-Gaussian2024.04 | 2.48 | — | — | — | |
| VDMType=Diff, Evaluation Protocol=variational bound, Data Augmentation=Extensive (A)2021.07 | 2.49 | — | — | — | |
| NFDMforward process parameterization=Gaussian2024.04 | 2.49 | — | — | — | |
| CR-NVAEType=VAE, Data Augmentation=Extensive (A)2021.07 | 2.51 | — | — | — | |
| Sparse Transformer + DistAugType=AR, Data Augmentation=Extensive (A)2021.07 | 2.53 | — | — | — | |
| MULANType=Diff2023.12 | 2.55 | — | — | — | |
| MuLAN2024.04 | 2.55 | — | — | — | |
| DistAug2020.12 | 2.56 | — | — | — | |
| DistAugPARAM=152.0M2021.06 | 2.56 | — | — | — | |
| i-DODE2024.04 | 2.56 | — | — | — | |
| VDMType=Diff, Evaluation Protocol=variational bound, Data Augmentation=None2021.07 | 2.65 | — | — | — | |
| VDM2024.04 | 2.65 | — | — | — | |
| Reflected Diffusion ModelsType=Diff2023.12 | 2.68 | — | — | — | |
| NDM2024.04 | 2.7 | — | — | — | |
| Sparse TransformerNumber of parameters=59M, Attention pattern=strided2019.04 | 2.8 | — | — | — | |
| Sparse Transformer2020.12 | 2.8 | — | — | — | |
| Sparse TransformerPARAM=59.0M2021.06 | 2.8 | — | — | — | |
| Sparse Transf.2D bias=2D sparse attn.2021.10 | 2.8 | — | — | — | |
| Sparse TransformerType=AR, Data Augmentation=None2021.07 | 2.8 | — | — | — | |
| ScoreFlowType=Diff, Evaluation Protocol=cont. norm. flow, Data Augmentation=Horizontal flips (C)2021.07 | 2.8 | — | — | — | |
| Our Decoder baselineModel Category=Autoregressive Models2019.01 | 2.83 | — | — | — | |
| ScoreFlowType=Diff, Evaluation Protocol=cont. norm. flow, Data Augmentation=None2021.07 | 2.83 | — | — | — | |
| ScoreFlowType=Diff2023.12 | 2.83 | — | — | — | |
| Score Flow2024.04 | 2.83 | — | — | — | |
| PixelSNAILModel Category=Autoregressive Models2019.01 | 2.85 | — | — | — | |
| PixelSNAIL2019.04 | 2.85 | — | — | — | |
| PixelSNAILArchitecture Category=Autoregressive2019.02 | 2.85 | — | — | — | |
| PixelSNAIL2D bias=2D conv. + attn.2021.10 | 2.85 | — | — | — | |
| S4 (large)2D bias=None2021.10 | 2.85 | — | — | — | |
| Very Deep VAEType=VAE, Data Augmentation=None2021.07 | 2.87 | — | — | — | |
| LSGMType=Diff, Data Augmentation=None2021.07 | 2.87 | — | — | — | |
| ImageTransformerModel Category=Autoregressive Models2019.01 | 2.9 | — | — | — | |
| Image Transformer2019.04 | 2.9 | — | — | — | |
| Image Transf.2D bias=2D local attn.2021.10 | 2.9 | — | — | — | |
| Image TransformerType=AR, Data Augmentation=None2021.07 | 2.9 | — | — | — | |
| ScoreFlowType=Diff, Evaluation Protocol=variational bound, Data Augmentation=Horizontal flips (C)2021.07 | 2.9 | — | — | — | |
| Image TransformerType=AR2023.12 | 2.9 | — | — | — | |
| NVAETIME (GPU hours)=552021.06 | 2.91 | — | — | — | |
| NVAEType=VAE, Data Augmentation=None2021.07 | 2.91 | — | — | — | |
| PixelCNN++Model Category=Autoregressive Models2019.01 | 2.92 | — | — | — | |
| PixelCNN++2019.04 | 2.92 | — | — | — | |
| PixelCNN++Architecture Category=Autoregressive2019.02 | 2.92 | — | — | — | |
| PixelCNN++2D bias=2D conv.2021.10 | 2.92 | — | — | — | |
| S4 (base)2D bias=None2021.10 | 2.92 | — | — | — | |
| PixelCNN++Type=AR, Data Augmentation=None2021.07 | 2.92 | — | — | — | |
| Improved DDPMType=Diff, Data Augmentation=None2021.07 | 2.94 | — | — | — | |
| Improved DDPM2024.04 | 2.94 | — | — | — | |
| DenseNet VLAEModel Category=Latent Variable Models2019.01 | 2.95 | — | — | — | |
| MAEArchitecture Category=Autoregressive2019.02 | 2.95 | — | — | — | |
| DenseFlowTIME (GPU hours)=2502021.06 | 2.98 | — | — | — | |
| Score SDEType=Diff, Data Augmentation=None2021.07 | 2.99 | — | — | — | |
| Flow MatchingType=Flow2023.12 | 2.99 | — | — | — | |
| Score SDE2024.04 | 2.99 | — | — | — | |
| Flow Matching2024.04 | 2.99 | — | — | — | |
| Stochastic Interp.2024.04 | 2.99 | — | — | — | |
| PixelRNN2016.05 | 3 | — | — | — | |
| PixelRNNModel Category=Autoregressive Models2019.01 | 3 | — | — | — | |
| PixelRNNArchitecture Category=Autoregressive2019.02 | 3 | — | — | — | |
| Row PixelRNN2D bias=2D BiLSTM2021.10 | 3 | — | — | — | |
| Gated PixelCNNModel Category=Autoregressive Models2019.01 | 3.03 | — | — | — | |
| PixelCNN2019.04 | 3.03 | — | — | — | |
| Gated PixelCNNArchitecture Category=Autoregressive2019.02 | 3.03 | — | — | — | |
| GPixelCNN2020.12 | 3.03 | — | — | — | |
| Gated PixelCNN2021.06 | 3.03 | — | — | — | |
| PixelCNNType=AR, Data Augmentation=None2021.07 | 3.03 | — | — | — | |
| PixelCNNType=AR2023.12 | 3.03 | — | — | — | |
| Flow++2020.12 | 3.08 | — | — | — | |
| Flow++PARAM=31.4M2021.06 | 3.08 | — | — | — | |
| Flow++Type=Flow, Data Augmentation=None2021.07 | 3.08 | — | — | — | |
| Flow++Architecture Category=Flow-based, Dequantization Strategy=Variational2019.02 | 3.09 | — | — | — | |
| Flow++variational dequantization=true2019.06 | 3.09 | — | — | — | |
| IAF VAEArchitecture Category=Autoregressive2019.02 | 3.11 | — | — | — | |
| ResNet VAE with IAFType=VAE, Data Augmentation=None2021.07 | 3.11 | — | — | — | |
| PixelCNN2D bias=2D conv.2021.10 | 3.14 | — | — | — | |
| MACOWArchitecture Category=Flow-based, Model Variation=+var, Dequantization Strategy=Variational2019.02 | 3.16 | — | — | — | |
| MaCowPARAM=43.5M2021.06 | 3.16 | — | — | — | |
| Diffusion Recovery Likelihood (T1k)Estimation Method=Approximated log partition function2020.12 | 3.18 | — | — | — | |
| EBM-DRLType=Diff, Data Augmentation=Horizontal flips (C)2021.07 | 3.18 | — | — | — | |
| MACOWArchitecture Category=Flow-based, Model Variation=+fine-grained2019.02 | 3.28 | — | — | — | |
| Residual Flow2019.06 | 3.28 | — | — | — | |
| Flow++Architecture Category=Flow-based, Dequantization Strategy=Uniform2019.02 | 3.29 | — | — | — | |
| Flow++variational dequantization=false2019.06 | 3.29 | — | — | — | |
| MACOWArchitecture Category=Flow-based, Model Variation=Original (Org)2019.02 | 3.31 | — | — | — | |
| EmergingEpochs=500, Convolution Type=Emerging convolutions2025.12 | 3.34 | — | — | — | |
| QuadEpochs=500, Convolution Type=3x3 invertible convolutions, Coupling Type=Quad-coupling2025.12 | 3.3471 | — | — | — | |
| 3 x 3Epochs=500, Convolution Type=3x3 invertible convolutions2025.12 | 3.3498 | — | — | — | |
| GlowArchitecture=multi-scale convolutional architectures2018.10 | 3.35 | — | — | — | |
| Glow2018.11 | 3.35 | — | — | — | |
| GlowArchitecture Category=Flow-based2019.02 | 3.35 | — | — | — | |
| Glow2019.06 | 3.35 | — | — | — | |
| Glow2020.12 | 3.35 | — | — | — | |
| GlowPARAM=44.0M2021.06 | 3.35 | — | — | — | |
| GLOW2022.05 | 3.35 | — | — | — | |
| GlowType=Flow, Data Augmentation=Small translations (B)2021.07 | 3.35 | — | — | — | |
| GlowEpochs=5002025.12 | 3.35 | — | — | — | |
| RNODEPARAM=1.4M, TIME (GPU hours)=31.82021.06 | 3.38 | — | — | — | |
| NSF2022.05 | 3.38 | — | — | — | |
| RNODE + STEERPARAM=1.4M, TIME (GPU hours)=22.22021.06 | 3.397 | — | — | — | |
| FFJORDArchitecture=multi-scale convolutional architectures2018.10 | 3.4 | — | — | — | |
| FFJORD2018.11 | 3.4 | — | — | — |