On the Anisotropy of Score-Based Generative Models

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Floros, Andreas, Moosavi-Dezfooli, Seyed-Mohsen, Dragotti, Pier Luigi
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866915579558887424
author Floros, Andreas
Moosavi-Dezfooli, Seyed-Mohsen
Dragotti, Pier Luigi
author_facet Floros, Andreas
Moosavi-Dezfooli, Seyed-Mohsen
Dragotti, Pier Luigi
contents We investigate the role of network architecture in shaping the inductive biases of modern score-based generative models. To this end, we introduce the Score Anisotropy Directions (SADs), architecture-dependent directions that reveal how different networks preferentially capture data structure. Our analysis shows that SADs form adaptive bases aligned with the architecture's output geometry, providing a principled way to predict generalization ability in score models prior to training. Through both synthetic data and standard image benchmarks, we demonstrate that SADs reliably capture fine-grained model behavior and correlate with downstream performance, as measured by Wasserstein metrics. Our work offers a new lens for explaining and predicting directional biases of generative models.
format Preprint
id arxiv_https___arxiv_org_abs_2510_22899
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle On the Anisotropy of Score-Based Generative Models
Floros, Andreas
Moosavi-Dezfooli, Seyed-Mohsen
Dragotti, Pier Luigi
Machine Learning
We investigate the role of network architecture in shaping the inductive biases of modern score-based generative models. To this end, we introduce the Score Anisotropy Directions (SADs), architecture-dependent directions that reveal how different networks preferentially capture data structure. Our analysis shows that SADs form adaptive bases aligned with the architecture's output geometry, providing a principled way to predict generalization ability in score models prior to training. Through both synthetic data and standard image benchmarks, we demonstrate that SADs reliably capture fine-grained model behavior and correlate with downstream performance, as measured by Wasserstein metrics. Our work offers a new lens for explaining and predicting directional biases of generative models.
title On the Anisotropy of Score-Based Generative Models
topic Machine Learning
url https://arxiv.org/abs/2510.22899