Postdoc, University of Oxford
1 paper at NeurIPS 2025
We show that diffusion models adapt to low-dimensional data geometry as a result of how approximations are formed at the level of the score function. We show that smoothing the score function naturally produces smoothing along the data manifold.
