Few-Shot Image Generation via Cross-Domain Correspondence

Resulting generator GAN adaptation …to paintings …to babies

Figure 1: Given a model trained on a large source dataset (G s ), we propose to adapt it to arbitrary image domains, so that the resulting model (G s→t ) captures these target distributions using extremely few training samples. In the process, our method discovers a one-to-one relation between the distributions, where noise vectors map to corresponding images in the source and target. Consequently, one can imagine how a natural face would look if Amedeo Modigliani had painted it, or how the face would look in its baby form. Please see our webpage for more results.