Reproduce “Diffusion-free Score Matching for Exact Conditional Sampling” 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
Thomas Wedenig, Robert Peharz
DISCO trains one score field for a slightly perturbed data distribution, rather than separate scores at each noise level. Fixing observed variables in the joint score yields the learned model’s exact conditional score, enabling direct conditional sampling. In two-dimensional tests, its samples more closely match each model’s reference conditional than tested diffusion samplers; this does not establish exactness for real data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.