SoLAD: Sampling Over Latent Adapter for Few Shot Generation

Abstract

Few-shot adaptation of Generative Adversarial Networks (GANs) under distributional shift is generally achieved via regularized retraining or latent space adaptation. While the former methods offer fast inference, the latter generate diverse images. This work aims to solve these issues and achieve the best of both regimes in a principled manner via Bayesian reformulation of the GAN objective. We highlight a hidden expectation term over GAN parameters, that is often overlooked but is critical in few-shot settings. This observation helps us justify prepending a latent adapter network (LAN) before a pre-trained GAN and propose a sampling procedure over the parameters of LAN (called SoLAD) to compute the usually-ignored hidden expectation. SoLAD enables fast generation of quality samples from multiple few-shot target domains using a GAN pre-trained on a single source domain.