Spectral Regularization for Diffusion Models

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Chandran, Satish, Santos, Nicolas Roque dos, Wu, Yunshu, Steeg, Greg Ver, Papalexakis, Evangelos
Format: Preprint
Veröffentlicht: 2026
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866912938841866240
author Chandran, Satish
Santos, Nicolas Roque dos
Wu, Yunshu
Steeg, Greg Ver
Papalexakis, Evangelos
author_facet Chandran, Satish
Santos, Nicolas Roque dos
Wu, Yunshu
Steeg, Greg Ver
Papalexakis, Evangelos
contents Diffusion models are typically trained using pointwise reconstruction objectives that are agnostic to the spectral and multi-scale structure of natural signals. We propose a loss-level spectral regularization framework that augments standard diffusion training with differentiable Fourier- and wavelet-domain losses, without modifying the diffusion process, model architecture, or sampling procedure. The proposed regularizers act as soft inductive biases that encourage appropriate frequency balance and coherent multi-scale structure in generated samples. Our approach is compatible with DDPM, DDIM, and EDM formulations and introduces negligible computational overhead. Experiments on image and audio generation demonstrate consistent improvements in sample quality, with the largest gains observed on higher-resolution, unconditional datasets where fine-scale structure is most challenging to model.
format Preprint
id arxiv_https___arxiv_org_abs_2603_02447
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Spectral Regularization for Diffusion Models
Chandran, Satish
Santos, Nicolas Roque dos
Wu, Yunshu
Steeg, Greg Ver
Papalexakis, Evangelos
Machine Learning
Diffusion models are typically trained using pointwise reconstruction objectives that are agnostic to the spectral and multi-scale structure of natural signals. We propose a loss-level spectral regularization framework that augments standard diffusion training with differentiable Fourier- and wavelet-domain losses, without modifying the diffusion process, model architecture, or sampling procedure. The proposed regularizers act as soft inductive biases that encourage appropriate frequency balance and coherent multi-scale structure in generated samples. Our approach is compatible with DDPM, DDIM, and EDM formulations and introduces negligible computational overhead. Experiments on image and audio generation demonstrate consistent improvements in sample quality, with the largest gains observed on higher-resolution, unconditional datasets where fine-scale structure is most challenging to model.
title Spectral Regularization for Diffusion Models
topic Machine Learning
url https://arxiv.org/abs/2603.02447