Reproduce “Optimizing Hierarchical Image VAEs for Sample Quality” 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

Eric Luhman, Troy Luhman

The authors modify hierarchical variational autoencoders (VAEs) to improve the perceptual quality of images generated from their prior. They redistribute information across latent groups and train a Gaussian decoder to tolerate pixel-level error; they also test sampling-time guidance for class-conditioned generation. On unconditional CIFAR-10, both changes reduce FID from 40.66 to 20.82, while test negative log-likelihood worsens from 3.10 to 4.39 bits per dimension.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.