Few-shot Image Generation Using Discrete Content Representation

TL;DR

This paper introduces a novel method for few-shot image generation that uses discrete content representations and autoregressive modeling to improve diversity and fidelity of generated images for unseen categories.

Contribution

It adapts few-shot image translation techniques to generation by quantizing content maps into discrete vectors and modeling their distribution conditioned on style vectors.

Findings

Produces higher diversity images than previous methods.

Achieves better image fidelity on real datasets.

Effectively handles unseen categories with limited data.

Abstract

Few-shot image generation and few-shot image translation are two related tasks, both of which aim to generate new images for an unseen category with only a few images. In this work, we make the first attempt to adapt few-shot image translation method to few-shot image generation task. Few-shot image translation disentangles an image into style vector and content map. An unseen style vector can be combined with different seen content maps to produce different images. However, it needs to store seen images to provide content maps and the unseen style vector may be incompatible with seen content maps. To adapt it to few-shot image generation task, we learn a compact dictionary of local content vectors via quantizing continuous content maps into discrete content maps instead of storing seen images. Furthermore, we model the autoregressive distribution of discrete content map conditioned on style vector, which can alleviate the incompatibility between content map and style vector. Qualitative and quantitative results on three real datasets demonstrate that our model can produce images of higher diversity and fidelity for unseen categories than previous methods.