We introduce label projection and embedding methods to improve generative adversarial networks. As proof of concept, we generate images of plasmonic nanostructures with different shapes by using a model trained on their linear optical response.