Image Generation on ImageNet 64x64

Metric: Bits per dim (higher is better)

LeaderboardDataset

Loading chart...

Results

Sort:

#	Model↕	Bits per dim▼	Extra Data	Paper	Date↕	Code
1	Logsparse (6 layers)	4.351	No	Enhancing the Locality and Breaking the Memory B...	2019-06-29	Code
2	Axial Transformer (6 layers)	4.032	No	Axial Attention in Multidimensional Transformers	2019-12-20	Code
3	Glow (Kingma and Dhariwal, 2018)	3.81	No	Glow: Generative Flow with Invertible 1x1 Convol...	2018-07-09	Code
4	Residual Flow	3.757	No	Residual Flows for Invertible Generative Modeling	2019-06-06	Code
5	MaCow (Unf)	3.75	No	MaCow: Masked Convolutional Generative Flow	2019-02-12	Code
6	Reformer (6 layers)	3.74	No	Reformer: The Efficient Transformer	2020-01-13	Code
7	Performer (6 layers)	3.719	No	Rethinking Attention with Performers	2020-09-30	Code
8	MALI	3.71	No	MALI: A memory efficient and reverse accurate in...	2021-02-09	Code
9	Reformer (12 layers)	3.71	No	Reformer: The Efficient Transformer	2020-01-13	Code
10	Parallel Multiscale	3.7	No	Parallel Multiscale Autoregressive Density Estim...	2017-03-10	-
11	Flow++	3.69	No	Flow++: Improving Flow-Based Generative Models w...	2019-02-01	Code
12	MaCow (Var)	3.69	No	MaCow: Masked Convolutional Generative Flow	2019-02-12	Code
13	Performer (12 layers)	3.636	No	Rethinking Attention with Performers	2020-09-30	Code
14	PixelCNN	3.57	No	PixelCNN Models with Auxiliary Variables for Nat...	2016-12-24	-
15	Gated PixelCNN (van den Oord et al., [2016c])	3.57	No	Conditional Image Generation with PixelCNN Decod...	2016-06-16	Code
16	Improved DDPM	3.53	No	Improved Denoising Diffusion Probabilistic Models	2021-02-18	Code
17	SPN	3.52	No	Generating High Fidelity Images with Subscale Pi...	2018-12-04	-
18	Very Deep VAE	3.52	No	Very Deep VAEs Generalize Autoregressive Models ...	2020-11-20	Code
19	Combiner-Mixture	3.504	No	Combiner: Full Attention Transformer with Sparse...	2021-07-12	Code
20	Sparse Transformer 59M (strided)	3.44	No	Generating Long Sequences with Sparse Transformers	2019-04-23	Code
21	MRCNF	3.44	No	Multi-Resolution Continuous Normalizing Flows	2021-06-15	Code
22	Hourglass	3.44	No	Hierarchical Transformers Are More Efficient Lan...	2021-10-26	Code
23	Routing Transformer	3.43	No	Efficient Content-Based Sparse Attention with Ro...	2020-03-12	Code
24	Combiner-Axial	3.42	No	Combiner: Full Attention Transformer with Sparse...	2021-07-12	Code
25	VDM	3.4	No	Variational Diffusion Models	2021-07-01	Code
26	NDM	3.35	No	Neural Diffusion Models	2023-10-12	-
27	FM	3.31	No	Flow Matching for Generative Modeling	2022-10-06	Code
28	BSI	3.22	No	Generative Modeling with Bayesian Sample Inference	2025-02-11	Code
29	NFDM	3.2	No	Neural Flow Diffusion Models: Learnable Forward ...	2024-04-19	Code
30	TarFlow	2.99	No	Normalizing Flows are Capable Generative Models	2024-12-09	Code

#1Logsparse (6 layers)SOTA
4.351
Bits per dim· 2019-06-29
Enhancing the Locality and Breaking the Memory Bottleneck of Transformer on Time Series Forecasting Code
#2Axial Transformer (6 layers)
4.032
Bits per dim· 2019-12-20
Axial Attention in Multidimensional Transformers Code
#3Glow (Kingma and Dhariwal, 2018)SOTA
3.81
Bits per dim· 2018-07-09
Glow: Generative Flow with Invertible 1x1 Convolutions Code
#4Residual Flow
3.757
Bits per dim· 2019-06-06
Residual Flows for Invertible Generative Modeling Code
#5MaCow (Unf)
3.75
Bits per dim· 2019-02-12
MaCow: Masked Convolutional Generative Flow Code
#6Reformer (6 layers)
3.74
Bits per dim· 2020-01-13
Reformer: The Efficient Transformer Code
#7Performer (6 layers)
3.719
Bits per dim· 2020-09-30
Rethinking Attention with Performers Code
#8MALI
3.71
Bits per dim· 2021-02-09
MALI: A memory efficient and reverse accurate integrator for Neural ODEs Code
#9Reformer (12 layers)
3.71
Bits per dim· 2020-01-13
Reformer: The Efficient Transformer Code
#10Parallel MultiscaleSOTA
3.7
Bits per dim· 2017-03-10
Parallel Multiscale Autoregressive Density Estimation
#11Flow++
3.69
Bits per dim· 2019-02-01
Flow++: Improving Flow-Based Generative Models with Variational Dequantization and Architecture Design Code
#12MaCow (Var)
3.69
Bits per dim· 2019-02-12
MaCow: Masked Convolutional Generative Flow Code
#13Performer (12 layers)
3.636
Bits per dim· 2020-09-30
Rethinking Attention with Performers Code
#14PixelCNN
3.57
Bits per dim· 2016-12-24
PixelCNN Models with Auxiliary Variables for Natural Image Modeling
#15Gated PixelCNN (van den Oord et al., [2016c])SOTA
3.57
Bits per dim· 2016-06-16
Conditional Image Generation with PixelCNN Decoders Code
#16Improved DDPM
3.53
Bits per dim· 2021-02-18
Improved Denoising Diffusion Probabilistic Models Code
#17SPN
3.52
Bits per dim· 2018-12-04
Generating High Fidelity Images with Subscale Pixel Networks and Multidimensional Upscaling
#18Very Deep VAE
3.52
Bits per dim· 2020-11-20
Very Deep VAEs Generalize Autoregressive Models and Can Outperform Them on Images Code
#19Combiner-Mixture
3.504
Bits per dim· 2021-07-12
Combiner: Full Attention Transformer with Sparse Computation Cost Code
#20Sparse Transformer 59M (strided)
3.44
Bits per dim· 2019-04-23
Generating Long Sequences with Sparse Transformers Code
#21MRCNF
3.44
Bits per dim· 2021-06-15
Multi-Resolution Continuous Normalizing Flows Code
#22Hourglass
3.44
Bits per dim· 2021-10-26
Hierarchical Transformers Are More Efficient Language Models Code
#23Routing Transformer
3.43
Bits per dim· 2020-03-12
Efficient Content-Based Sparse Attention with Routing Transformers Code
#24Combiner-Axial
3.42
Bits per dim· 2021-07-12
Combiner: Full Attention Transformer with Sparse Computation Cost Code
#25VDM
3.4
Bits per dim· 2021-07-01
Variational Diffusion Models Code
#26NDM
3.35
Bits per dim· 2023-10-12
Neural Diffusion Models
#27FM
3.31
Bits per dim· 2022-10-06
Flow Matching for Generative Modeling Code
#28BSI
3.22
Bits per dim· 2025-02-11
Generative Modeling with Bayesian Sample Inference Code
#29NFDM
3.2
Bits per dim· 2024-04-19
Neural Flow Diffusion Models: Learnable Forward Process for Improved Diffusion Modelling Code
#30TarFlow
2.99
Bits per dim· 2024-12-09
Normalizing Flows are Capable Generative Models Code