Search papers, labs, and topics across Lattice.
University of Cambridge
3
0
5
A learned diffusion model can achieve a perfect match to the target distribution while exhibiting an arbitrarily large $L^2$ score error, challenging existing training paradigms.
Unlock probabilistic PDE modeling from existing deterministic backbones with a simple retrofitting technique that slashes error by up to 54% and works across diverse architectures.
Diffusion models degrade predictably when trained on their own outputs, but a little fresh data can keep them on track.
