GAMEX seminar series
Description: This online seminar is organized within the GAMEX project with the support of the Glasgow–Edinburgh Extremes Network (GLE²N) and CIRCE.
Date: 2 October 2026
Time: 15:30 CEST (13:30 UTC / 14:30 BST)
Speaker: Antonio Ocello (ENSAE Paris)
Title:
Do Heavy Tails Help Diffusion? On the Subtle Trade-off Between Initialization and Training
Abstract:
Recent works have proposed incorporating heavy-tailed (HT) noise into diffusion- and flow-based generative models, with the goals of better recovering the tails of target distributions and improving generative diversity. This motivation is intuitive: if the data are heavy-tailed, HT noise may appear better matched than light-tailed (LT) Gaussian noise. However, replacing Gaussian noise by HT noise also changes the underlying estimation problem.
In this talk, we revisit this paradigm through a combined theoretical and empirical study, establishing sampling-error bounds for two representative diffusion models driven by HT and LT noise. We show that HT noise makes the statistical estimation problem harder, leading to less favorable sampling-error bounds. We support these findings with experiments on synthetic and real-world datasets, empirically recovering the predicted error trade-off.
Our results call into question a growing design trend in generative modeling and challenge the use of HT noise to improve rare-region exploration.
This talk is based on joint work with Hamza Cherkaoui and Hélène Halconruy:
arXiv:2605.13175
Access link: Join the seminar on Microsoft Teams
Meeting ID: 349 080 320 615 578
Passcode: ZL64qq2j