Diffusion models under low-noise regime

Elizabeth Pavlova, Xue-Xin Wei · 2025

Read Diffusion models under low-noise regime by Elizabeth Pavlova, Xue-Xin Wei (2025) 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-2506-07841.txt.

Description

Elizabeth Pavlova and Xue-Xin Wei probe diffusion models in the gentle-noise zone in this 2025 paper. These image generators learn by cleaning up corrupted pictures. Recent work split their behavior into memorizing training pictures and inventing new ones, but mostly at heavy noise. What they do as denoisers when the corruption is faint was murky.

The team tested models on celebrity face subsets of varied sizes and on tidy math benchmarks. Models trained on separate data sets agreed at high noise yet diverged near the clean data, each hugging its own training set. Training size, data shape, and the training goal all steered the cleaning path and the accuracy of the learned score. The work matters because real uses often start from slightly flawed inputs.

Catalog record

CREATORElizabeth Pavlova, Xue-Xin Wei
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE39 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2506-07841

Browse the full library · About the Tower of Babel