The Interplay of Data Structure and Imbalance in the Learning Dynamics of Diffusion Models

Flavio Nicoletti, Chenxiao Ma, Enrico Ventura, Luca Saglietti, Stefano Sarao Mannelli · 2026

Read The Interplay of Data Structure and Imbalance in the Learning Dynamics of Diffusion Models by Flavio Nicoletti, Chenxiao Ma, Enrico Ventura, Luca Saglietti, Stefano Sarao Mannelli (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-2605-06367.txt.

Description

Flavio Nicoletti and four co-authors examine how data structure and class imbalance shape the learning dynamics of diffusion models. Real datasets mix classes with different spreads and frequencies. Models usually move from generalizing to memorizing, but older theory assumes uniform data, leaving open how uneven classes change that order.

Using a random-features model trained on Gaussian mixtures, the authors derive which classes generalize or memorize first. Class variance matters most, centroid geometry second, and sampling imbalance can reverse the order or delay minority classes. Tests with U-Net models on Fashion MNIST support the predictions, warning that some classes may memorize while others stay underlearned.

Catalog record

CREATORFlavio Nicoletti, Chenxiao Ma, Enrico Ventura, Luca Saglietti, Stefano Sarao Mannelli
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE147 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.06367

Browse the full library · About the Tower of Babel