The Euclidean Space is Evil: Hyperbolic Attribute Editing for Few-shot Image Generation

Lingxiao Li,Yi Zhang,Shuhui Wang

2023-08-19

Abstract:Few-shot image generation is a challenging task since it aims to generate diverse new images for an unseen category with only a few images. Existing methods suffer from the trade-off between the quality and diversity of generated images. To tackle this problem, we propose Hyperbolic Attribute Editing~(HAE), a simple yet effective method. Unlike other methods that work in Euclidean space, HAE captures the hierarchy among images using data from seen categories in hyperbolic space. Given a well-trained HAE, images of unseen categories can be generated by moving the latent code of a given image toward any meaningful directions in the Poincaré disk with a fixing radius. Most importantly, the hyperbolic space allows us to control the semantic diversity of the generated images by setting different radii in the disk. Extensive experiments and visualizations demonstrate that HAE is capable of not only generating images with promising quality and diversity using limited data but achieving a highly controllable and interpretable editing process.

Computer Vision and Pattern Recognition

What problem does this paper attempt to address?

The paper attempts to address the problem of generating high-quality and diverse images with only a small number of samples, known as the few-shot image generation problem. Specifically, existing methods often face a trade-off between quality and diversity when generating new images. This challenge is particularly pronounced when dealing with unseen categories, where generating new, realistic, and diverse images based on only a few samples is a significant challenge.

To overcome this difficulty, the authors propose a new method called "Hyperbolic Attribute Editing" (HAE). Unlike traditional methods, HAE utilizes hyperbolic space instead of Euclidean space to capture the hierarchical relationships between images. Through this approach, not only can the quality and diversity of generated images be improved, but a highly controllable and interpretable editing process can also be achieved. The core of HAE lies in mapping the latent vectors of images from Euclidean space to hyperbolic space and controlling the semantic diversity of generated images by adjusting the position of latent codes in hyperbolic space. This method is particularly suitable for few-shot image generation tasks because it can leverage data from known categories to guide the generation of images from unknown categories while maintaining the diversity and quality of the generated images.