Reproduce “Binary Latent Diffusion” 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

Ze Wang, Jiang Wang, Zicheng Liu, Qiang Qiu

Binary Latent Diffusion presents a framework that uses a binary latent space and a Bernoulli diffusion process to generate high-quality, high-resolution images. This approach achieves competitive performance with dramatically fewer sampling steps than continuous latent diffusion models.

Key insights

  • A binary latent space can provide both compact and expressive image representations, enabling higher quality without large codebooks or multi-stage architectures.
  • A Bernoulli diffusion process, specifically designed for discrete binary states, dramatically improves sampling efficiency, allowing high-quality generation with significantly fewer steps.
  • Reparameterizing the reverse diffusion process to predict the "flipping probability" of binary codes enhances training stability and overall generative performance.

Summary by alphaXiv.

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