Multiple one-shot image generation via deep structure reshuffle.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• One-Shot Image Generation (OSG) models require minimal training data but struggle with multi-class generation.
• Existing OSG methods are limited to modeling single classes, hindering broader applications.

Purpose of the Study:

• To introduce a unified framework, Multiple One-Shot Image Generation (MultiOSG), for generating images from multiple classes simultaneously.
• To address the limitations of classical OSG methods in handling diverse sample classes.

Main Methods:

• Propose the MultiOSG method, a unified framework for generating multiple one-shot images.
• Disentangle images into structure and texture codes.
• Introduce Structure Reshuffle to obtain new structure codes with local diversity and global controllability.

Main Results:

• MultiOSG demonstrates competitive performance in one-shot image generation tasks.
• The method successfully models multiple images and generates random samples across various classes.
• Experimental results validate the effectiveness and extended capabilities of MultiOSG.

Conclusions:

• MultiOSG effectively overcomes the multi-class limitation of traditional OSG.
• The proposed framework expands the applicability of one-shot image generation.
• The method offers a novel approach to generating diverse image sets from limited data.