Glow: Generative Flow with Invertible 1x1 Convolutions

Flow-based generative models (Dinh et al., 2014) are conceptually attractive due to tractability of the exact log-likelihood, tractability of exact latent-variable inference, and parallelizability of both training and synthesis. In this paper we propose Glow, a simple type of generative flow using an invertible 1x1 convolution... (read more)

PDF Abstract NeurIPS 2018 PDF NeurIPS 2018 Abstract
TASK DATASET MODEL METRIC NAME METRIC VALUE GLOBAL RANK RESULT BENCHMARK
Image Generation CelebA 256x256 Glow (Kingma and Dhariwal, 2018) bpd 1.03 # 8
Image Generation CIFAR-10 Glow (Kingma and Dhariwal, 2018) bits/dimension 3.35 # 23
Image Generation ImageNet 32x32 Glow (Kingma and Dhariwal, 2018) bpd 4.09 # 13
Image Generation ImageNet 64x64 Glow (Kingma and Dhariwal, 2018) Bits per dim 3.81 # 13

Results from Other Papers


TASK DATASET MODEL METRIC NAME METRIC VALUE GLOBAL RANK SOURCE PAPER COMPARE
Image Generation CelebA-HQ 256x256 GLOW FID 68.93 # 4

Methods used in the Paper


METHOD TYPE
Invertible 1x1 Convolution
Convolutions
Affine Coupling
Bijective Transformation
Normalizing Flows
Distribution Approximation
1x1 Convolution
Convolutions
Adam
Stochastic Optimization
ReLU
Activation Functions
Convolution
Convolutions
Activation Normalization
Normalization
GLOW
Generative Models