Reproduce “Gotta Go Fast When Generating Data with Score-Based 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
Alexia Jolicoeur-Martineau, Ke Li, Rémi Piché-Taillefer, Tal Kachman, Ioannis Mitliagkas
The authors introduce an adaptive solver for continuous-time image generation that adjusts reverse-step sizes and compares two estimates to decide whether a step is acceptable, while using the higher-order extrapolated estimate when it meets a chosen tolerance. On 256×256 FFHQ, it achieves FID 15.67 with 643 score-network evaluations, versus 44.42 for Euler–Maruyama at the same budget; results are limited to tested image models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.