Initialization-Aware Score-Based Diffusion Sampling

Tiziano Fassina, Gabriel Cardoso, Sylvan Le Corff, Thomas Romary · 2026

Read Initialization-Aware Score-Based Diffusion Sampling by Tiziano Fassina, Gabriel Cardoso, Sylvan Le Corff, Thomas Romary (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-00772.txt.

Description

Fassina and three colleagues extend score based generative models to heavy tailed distributions, a largely open regime where dedicated models lack clear quality and solid theory. First, early stopping plus suitable initialisation is enough to extend diffusion to any target, with soundness of the backward process and convergence in KL divergence. Second, normalising flows gain convergence results under mild flow family conditions with no tail assumptions on the target. A unified framework then trains a flow to capture tail behaviour first and uses it as the diffusion starting point.

Catalog record

CREATORTiziano Fassina, Gabriel Cardoso, Sylvan Le Corff, Thomas Romary
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE107293
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.00772

Browse the full library · About the Tower of Babel