Reproduce “Diffusion Models under Alternative Noise: Simplified Analysis and Sensitivity” 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

Juhyeok Choi, Chenglin Fan

Researchers at Seoul National University simplified the theoretical analysis of diffusion model sampling, demonstrating that computationally inexpensive discrete noise distributions can replace standard Gaussian noise without compromising strong convergence rates. This holds true as long as the alternative noise maintains zero mean and unit variance, leading to potential efficiencies in hardware deployment.

Key insights

  • The strong convergence rate of Euler-Maruyama discretization for VP-SDEs, $O(d^{1/2}T^{-1/2})$, is preserved when Gaussian noise is replaced by alternative discrete noise, provided its first two moments are matched.
  • The robust generative capabilities of diffusion models are primarily driven by the first two moments (mean and variance) of the noise process, not its specific distribution shape.
  • Symmetry of the noise distribution is crucial for maintaining sample quality, as asymmetric distributions like Laplace noise significantly degrade performance.

Summary by alphaXiv.

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