8000 GitHub - mk-0/minincsn: Noise-conditional score network in 200 lines of code · GitHub
[go: up one dir, main page]

Skip to content
mk-0Public

About

Noise-conditional score network in 200 lines of code

Resources

Stars

0 stars

Watchers

1 watching

Forks

Latest commit

 

History

1 Commit

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 

Repository files navigation

Noise Conditional Score Network in 200 lines of code

MNIST digits sampled from NCSN

NCSN is a precursor to score-based DDPM (diffusion models). The difference is that instead of reversing a single step of the process of partial data corruption, it attempts to reconstruct the original image directly from any noise level. This makes it easy to express in code and easy to understand.

But because the training objective is different from the sampling procedure (which is still the same as in diffusion models), NCSN models are unstable and hard to scale to high resolution images. Diffusion models address this gap.

Nevertheless, a simple implementation like this one is helpful to understand diffusion models and generative image models in general.

Sources:

  1. Song Y. and Ermon S. Generative Modeling by Estimating Gradients of the Data Distribution, 2019
  2. Song Y. and Ermon S. Improved Techniques for Training Score-Based Generative Models, 2020

About

Noise-conditional score network in 200 lines of code

Resources

Stars

0 stars

Watchers

1 watching

Forks

Releases

Packages

Contributors

Languages

0