Reproduce “Score-Based Generative Modeling through Stochastic Differential Equations” 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

Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, Ben Poole

The authors introduce a continuous-time framework for generating images by reversing a stochastic process from noise toward data. A neural network estimates noisy-data scores to guide reverse sampling, supporting predictor-corrector methods and conditional generation. On unconditional CIFAR-10, their VE model reports FID 2.20 and Inception Score 9.89, versus DDPM’s 3.17 and 9.46; these results apply to the reported setup.

Key insights

  • Score-based generative models, including SMLD and DDPM, can be unified as discrete approximations of continuous-time Stochastic Differential Equations (SDEs).
  • Generative sampling can be framed as solving a reverse-time SDE, allowing for flexible numerical solvers and advanced techniques like Predictor-Corrector samplers.
  • A deterministic probability flow Ordinary Differential Equation (ODE) exists with the same marginals as the SDE, enabling exact likelihood computation and uniquely identifiable data encodings.

Summary by alphaXiv.

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