Reproduce “HiLLoC: Lossless Image Compression with Hierarchical Latent Variable 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
James Townsend, Thomas Bird, Julius Kunze, David Barber
University College London researchers present HiLLoC, a framework that scales VAE-based lossless image compression to large color photographs using hierarchical latent variable models and a vectorized Asymmetric Numeral Systems implementation. HiLLoC achieved a compression rate of 3.15 bits/dimension on full-size ImageNet, outperforming established codecs and deep learning methods, while operating with significantly improved computational efficiency.
Key insights
- Fully convolutional VAEs exhibit strong generalization capabilities, allowing models trained on small images to effectively compress much larger, higher-resolution images without modification.
- Vectorization of arithmetic coding operations significantly boosts computational efficiency, making deep generative model-based lossless compression practical for large-scale data.
- Dynamic discretization based on conditional priors simplifies the integration of continuous hierarchical latent variables into discrete entropy coding schemes like ANS, enabling top-down inference.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.