I was in a Zoom meeting when someone called a paper published at NeurIPS 2023 “old.”😅

In the interim, I wanted to advertise our YouTube channel - youtube.com/@montecarlos... - which contains recordings for the bulk of our talks so far (sites.google.com/view/monte-c..., sites.google.com/view/monte-c...). I encourage you to catch up and enjoy them over the intervening months!

Yeah, my guess is that generating pretty images does not necessarily require accurate sampling from the original distribution. That said, increasing the number of discretization steps still improves metrics like FID for DDIM.

If you think the process as targeting at p(x) using proposal q(x, x’) = \int_y p(y|x) p(x’|y) dy, then it should have acceptance 1?

Honorable mentions

Can we directly use Ascolani-Lavenant-Zanella on this? I thought their entropy contraction implies spectral gap, spectral gap implies mixing time under warm start.