Multiple one-shot image generation via deep structure reshuffle.

One-Shot image Generation (OSG) models have attracted much attention in recent years due to their relaxed requirements on the number of training samples (one is required). However, these methods cannot simultaneously model multiple samples of different classes (e.g., "balloons" and "birds" images), which limits the application and development of the models. To this end, we propose a unified framework to solve this problem, called Multiple One-Shot image Generation (MultiOSG) method. The core design of MultiOSG is to disentangle the image into a structure code and a texture code and obtain a new structure code through the proposed Structure Reshuffle. Compared with the original structure code, the new one should have good local diversity and global controllability. Then, the new structure and original texture codes are combined to obtain the corresponding generated results. Qualitative and quantitative experimental results show that our MultiOSG is competitive in the OSG tasks. Moreover, MultiOSG can model multiple images and produce random samples of various classes, which extends the limitation of the classical OSG method. Code will be released upon acceptance at https://github.com/gouayao/MultiOSG.