Few Shot Image Generation Using Conditional Set-Based GANs
Abstract: While there have been tremendous advances made in few-shot and zero-shot image generation in recent years, one area that remains comparatively underexplored is few-shot generation of images conditioned on sets of unseen images. Existing methods typically condition on a single image only and require strong assumptions about the similarity of the latent distribution of unseen classes relative to training classes. In contrast, we propose SetGAN - a conditional, set-based GAN that learns to generate sets of images conditioned on reference sets from unseen classes. SetGAN can combine information from multiple reference images, as well as generate diverse sets of images which mimic the factors of variation within the reference class. We also identify limitations of existing performance metrics for few-shot image generation, and discuss alternative performance metrics that can mitigate these problems.
Submission Length: Long submission (more than 12 pages of main content)
Assigned Action Editor: ~Abhishek_Kumar1
Submission Number: 1922
Loading