Attributed Synthetic Data Generation for Zero-shot Image Classification

Shijian Wang, Linxin Song, Ryotaro Shimizu, Masayuki Goto, Hanqian wu

Keywords: Synthetic image, Zero-shot image classification, Diffusion

Abstract: Zero-shot image classification is a challenging task aiming to classify real images without real training examples. Recent research has employed synthetic training images generated by text-to-image models to address the challenge. However, existing approaches heavily rely on simplistic prompt strategies, which limit the diversity of the synthetic images. In this paper, we propose AttrSyn, which leverages large language models to obtain attributed prompts. These prompts allow for the generation of more diverse attributed images (e.g., specifying attributes such as style and background). By conducting experiments on two fine-grained datasets, we demonstrate that AttrSyn significantly outperforms simple base prompts, regardless of the visual encoder and classifier settings.

Submission Number: 13