Reproduce “Wavelet Flow: Fast Training of High Resolution Normalizing Flows” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jason J. Yu, Konstantinos G. Derpanis, Marcus A. Brubaker
Wavelet Flow presents a multi-scale normalizing flow architecture built on wavelet decompositions, which enables training speedups of up to 15x over previous methods. This approach facilitates the first successful training of normalizing flows on 1024x1024 pixel images, yielding competitive density estimation performance.
Key insights
- Orthonormal wavelet transforms have a Jacobian determinant of 1, meaning the probability density in the image space equals the density in the wavelet coefficient space.
- The multi-scale nature of wavelets allows the image distribution to be factorized into independent components (coarse approximation and detail coefficients), enabling modular and parallel training.
- Conditioning higher-resolution detail flows on lower-resolution image approximations provides a natural mechanism for multi-scale generative modeling.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.