A Diffusion-Based Data Augmentation Approach for Synthetic Human Portraits Dataset

ICLR 2026 Conference Submission24525 Authors

Keywords: Diffusion Models; Data Augmentation; Transfer Learning; Image-to-Image Translation; Human Synthetic Dataset.

Abstract: Deep learning models have achieved remarkable success in computer vision. However, their generalizability remains limited when applied to new tasks. Data augmentation can help mitigate this issue, but traditional augmentation methods, such as rotation and scaling, are easy to conduct, but are also becoming increasingly inadequate when facing modern machine learning tasks. To address this issue, we propose a diffusion-based image-to-image augmentation workflow that transforms the original human images into new samples while keeping biometrical data unchanged, enriching the dataset without altering its key features. The resulted augmented dataset contains 225 synthetic anatomical models, each containing 44 images, resulting in a total of 9,900 images. Evaluation experiments demonstrate that the augmented dataset maintains 99.99% classification accuracy after 200 training epochs. Moreover, the quantitative evaluation shows that the maximum pixel deviation among all selected facial keypoints is only 10.1 pixels, with most remaining within 5–8 pixels, indicating that the new dataset is highly consistent with the original one. These findings demonstrate that diffusion-based augmentation is able to expand data diversity without compromising model performance nor the data accuracy.

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 24525