Abstract:
Training Generative Adversarial Networks (GANs) requires large amounts of data, and creating sufficiently large datasets to generate images of faces with attributes of a ...Show MoreMetadata
Abstract:
Training Generative Adversarial Networks (GANs) requires large amounts of data, and creating sufficiently large datasets to generate images of faces with attributes of a specific group is difficult. In this study, we propose a few-shot method to rapidly generate high-quality images of faces with features associated with a specific group. Our proposed approach comprises two techniques, Pivotal Tuning Inversion (PTI) and Principal Component Analysis (PCA). PTI consists of two processes: Inversion and Pivotal Tuning. The images of the dataset are projected into the latent space as latent vectors by Inversion, and the pretrained weights of the generator are tuned to adapt to the dataset by Pivotal Tuning. The distribution with the features of the group is obtained from the latent vectors using PCA. Since training on the dataset requires only a short computational time by using these techniques, our method is faster than the previous few-shot methods in the image generation process. With three datasets each containing approximately 100 face images of members of specific groups, the experimental results showed that the proposed method achieved the generation of higher-quality images with the features of the group in a shorter computational time than all previous few-shot methods.
Published in: 2023 International Conference on Artificial Intelligence in Information and Communication (ICAIIC)
Date of Conference: 20-23 February 2023
Date Added to IEEE Xplore: 23 March 2023
ISBN Information: