Exploiting mixing regularization for truly unsupervised font synthesis.

•An unsupervised generative adversarial network (GAN) architecture for font generation.•Mixing regularization is exploited for learning font content and style.•A projection encoder is jointly trained with GAN to learn the glyph projection to its latent space.•The proposed model performs competitive even compared against other supervised models•Can be used for other tasks such as font attribute control, word image generation, etc.