Reproduce “Denoising Deep Generative Models” 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
Gabriel Loaiza-Ganem, Brendan Leigh Ross, Luhuan Wu, John P. Cunningham, Jesse C. Cresswell, Anthony L. Caterini
The authors test whether adding Gaussian noise and denoising can help likelihood-based generative models address manifold overfitting, where high likelihood can coexist with a wrong distribution along the data manifold. They evaluate VAEs and normalizing flows on four image datasets. On MNIST, Tweedie-denoised NF has FID 105.6 ± 0.5 versus 103.2 ± 0.4 without denoising; gains are sporadic.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.