DataTerminology.comNEW
A flow-based generative model introducing invertible 1x1 convolutions for efficient, high-quality image synthesis.
DESCRIPTION
Glow is a normalizing flow model that extends RealNVP with invertible 1×1 convolutions and activation normalization (actnorm). These innovations improve channel mixing and training stability, enabling high-quality image generation and manipulation. Glow provides exact likelihoods and efficient sampling through invertible transformations with tractable Jacobians. It supports multi-scale architectures to handle large images. The model was released by OpenAI with code and visual results demonstrating realistic synthesis and latent-space editing.
ALTERNATIVE DEFINITIONS
arXiv
We present Glow, a generative flow with invertible 1x1 convolutions.
Highlights the defining invertible 1×1 convolution operation.
OpenAI
Glow is a generative model for image synthesis that uses invertible 1x1 convolutions to improve flow-based generative modeling.
Founder perspective emphasizing application to image synthesis.
Wikipedia
Normalizing flows are generative models constructed by a succession of invertible transformations with tractable Jacobians.
Places Glow within the broader normalizing flow framework.
RELATED TERMS
Normalizing Flow Invertible 1x1 Convolution ActNorm Variational Autoencoder GAN Latent Variables Change of Variables Kernel Density Estimation
COMPETING TERMS
RealNVP Masked Autoregressive Flow
KNOWLEDGE AREAS
TARGET AUDIENCE
AI Engineer Data Scientist MLOps Engineer
EXTERNAL SOURCES
- Wikipedia - Normalizing flow
- arXiv - Glow original paper
- OpenAI - Glow research overview
- Papers with Code - Paper summary and benchmarks
TOP QUESTIONS
What makes Glow different from RealNVP?
Glow introduces invertible 1×1 convolutions for channel mixing and actnorm for stable training, improving image modeling quality.
Category: Comparison
Why is Glow attractive for image synthesis?
It offers exact likelihoods, efficient sampling, and smooth latent interpolation enabling edits and realistic generation.
Category: Benefits
What is an invertible 1×1 convolution?
It’s a learned, invertible linear transform across channels at each spatial location, enabling efficient, bijective mixing.
Category: Concept
What are Glow’s limitations?
Memory usage can be high for large images, and it may lag GANs in sample fidelity for some domains.
Category: Limitations
How is training stability improved in Glow?
Activation normalization (actnorm) provides data-dependent initialization and per-channel scaling, stabilizing optimization.
Category: Technique
Where is Glow used?
Primarily in research for density estimation, controllable image generation, and as a baseline in flow-based modeling.
Category: Usage
How does Glow sample images?
By sampling from a simple base distribution and applying the inverse sequence of invertible transformations.
Category: Technique
Does Glow provide exact log-likelihoods?
Yes. Its invertible transformations yield tractable Jacobians for exact likelihood computation.
Category: Benefits
What architectural design supports high-resolution images in Glow?
A multi-scale architecture splits and factors out variables across levels to manage compute and memory.
Category: Architecture