Regularization for Unconditional Image Diffusion Models via Shifted Data Augmentation

Diffusion models are a powerful class of techniques in ML for generating realistic data, but they are highly prone to overfitting, especially with limited training data. While data augmentation such as image rotation can mitigate this issue, it often causes leakage, where augmented content appears i...

Full description

Saved in:
Bibliographic Details
Main Authors: Kensuke Nakamura, Bong-Soo Sohn, Simon Korman, Byung-Woo Hong
Format: Article
Language:English
Published: IEEE 2025-01-01
Series:IEEE Access
Subjects:
Online Access:https://ieeexplore.ieee.org/document/11048911/
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1849415661505216512
author Kensuke Nakamura
Bong-Soo Sohn
Simon Korman
Byung-Woo Hong
author_facet Kensuke Nakamura
Bong-Soo Sohn
Simon Korman
Byung-Woo Hong
author_sort Kensuke Nakamura
collection DOAJ
description Diffusion models are a powerful class of techniques in ML for generating realistic data, but they are highly prone to overfitting, especially with limited training data. While data augmentation such as image rotation can mitigate this issue, it often causes leakage, where augmented content appears in generated samples. In this paper, we propose a novel regularization framework, called shifted data-augmentation (SDA), for training unconditional diffusion models. SDA introduces an auxiliary diffusion path using transformed data and the noise-shift technique alongside the standard path with original data. This dual-path structure enables effective regularization while suppressing leakage through a trajectory shift in the diffusion process. We implement SDA with image rotation as its most basic and interpretable configuration. We also conduct synthetic and empirical analyses demonstrating that SDA effectively leverages the regularization benefit of image rotation. In particular, SDA yielded notable performance in training with limited data.
format Article
id doaj-art-13170482bc8d4e828c95c2aac97c3cb3
institution Kabale University
issn 2169-3536
language English
publishDate 2025-01-01
publisher IEEE
record_format Article
series IEEE Access
spelling doaj-art-13170482bc8d4e828c95c2aac97c3cb32025-08-20T03:33:27ZengIEEEIEEE Access2169-35362025-01-011311325811327310.1109/ACCESS.2025.358275611048911Regularization for Unconditional Image Diffusion Models via Shifted Data AugmentationKensuke Nakamura0https://orcid.org/0000-0002-6858-3551Bong-Soo Sohn1https://orcid.org/0000-0003-4656-5659Simon Korman2Byung-Woo Hong3https://orcid.org/0000-0003-2752-3939Artificial Intelligence Department, Chung-Ang University, Seoul, South KoreaComputer Science Department, Chung-Ang University, Seoul, South KoreaDepartment of Computer Science, University of Haifa, Haifa, IsraelArtificial Intelligence Department, Chung-Ang University, Seoul, South KoreaDiffusion models are a powerful class of techniques in ML for generating realistic data, but they are highly prone to overfitting, especially with limited training data. While data augmentation such as image rotation can mitigate this issue, it often causes leakage, where augmented content appears in generated samples. In this paper, we propose a novel regularization framework, called shifted data-augmentation (SDA), for training unconditional diffusion models. SDA introduces an auxiliary diffusion path using transformed data and the noise-shift technique alongside the standard path with original data. This dual-path structure enables effective regularization while suppressing leakage through a trajectory shift in the diffusion process. We implement SDA with image rotation as its most basic and interpretable configuration. We also conduct synthetic and empirical analyses demonstrating that SDA effectively leverages the regularization benefit of image rotation. In particular, SDA yielded notable performance in training with limited data.https://ieeexplore.ieee.org/document/11048911/Data augmentationdiffusion modelnoise-shiftregularization
spellingShingle Kensuke Nakamura
Bong-Soo Sohn
Simon Korman
Byung-Woo Hong
Regularization for Unconditional Image Diffusion Models via Shifted Data Augmentation
IEEE Access
Data augmentation
diffusion model
noise-shift
regularization
title Regularization for Unconditional Image Diffusion Models via Shifted Data Augmentation
title_full Regularization for Unconditional Image Diffusion Models via Shifted Data Augmentation
title_fullStr Regularization for Unconditional Image Diffusion Models via Shifted Data Augmentation
title_full_unstemmed Regularization for Unconditional Image Diffusion Models via Shifted Data Augmentation
title_short Regularization for Unconditional Image Diffusion Models via Shifted Data Augmentation
title_sort regularization for unconditional image diffusion models via shifted data augmentation
topic Data augmentation
diffusion model
noise-shift
regularization
url https://ieeexplore.ieee.org/document/11048911/
work_keys_str_mv AT kensukenakamura regularizationforunconditionalimagediffusionmodelsviashifteddataaugmentation
AT bongsoosohn regularizationforunconditionalimagediffusionmodelsviashifteddataaugmentation
AT simonkorman regularizationforunconditionalimagediffusionmodelsviashifteddataaugmentation
AT byungwoohong regularizationforunconditionalimagediffusionmodelsviashifteddataaugmentation