Research Director at CNRS in Computer Sciences in Lyon and part-time Assistant Professor at École Polytechnique, Paris.
Nov 20, 2023 • 14 tweets • 5 min read
You probably have a vague idea of how diffusion models work: progressively add noise to images during training, and learn a model to remove that noise. To generate a new image, just denoise a "pure noise" image. Let's provide some intuition to the formulas involved. đź”˝
First there is the basic idea that for training, you need to transform your image into some noise. It would be tempting to add a random value (e.g., gaussian) to each pixel to add noise. However, the goal is to completely destroy any signal.
May 5, 2022 • 17 tweets • 8 min read
Super happy to share our @siggraph paper with @nvidia "MatBuilder: Mastering Sampling Uniformity Over Projections" (@LoisResearch, @nbonneel, @dcoeurjo, JC.Iehl, A.Keller, V.Ostromoukhov) projet.liris.cnrs.fr/matbuilder/. The motivation is that to estimate the value of an integral [1/16]
we can randomly sample the integrand and sum values -- this is Monte Carlo integration. With "well spread" samples instead of naive random values, we can improve the accuracy of the integral estimate a lot -- this is the quasi-Monte Carlo method. [2/16]