LAMDA: Latent mapping for domain adaption of image generators

Yazeed Alharbi, Peter Wonka

Keywords: domain adaptation, gan, image synthesis, image generative, generative models

Abstract: Our paper tackles the problem of adapting image generators to new keyword-defined domains without training on any new images. We combine the power of CLIP models for image-text similarity with the disentangled representation of images found in the latent spaces of generative adversarial networks (GANs). We present the latent mapper (LAMDA) which maps directions in CLIP text space to directions in the GAN latent space. Using a latent mapper enables training on a large number of keywords simultaneously which was not previously possible, and allows benefiting from the inter-relation of different keywords. It also leads to higher image quality while requiring only a fraction of the training time and parameters of state-of-the-art methods. Our method allows the network to learn the relationship between words and latent spaces. As a result it allows the composition of words to generate semantically meaningful images. Additionally, it allows adaptation to some unseen words after training. We perform extensive analysis to validate the advantages of our method quantitatively and qualitatively.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Generative models

TL;DR: We adapt GANs to new domains without training on new images. This is done by only learning how to translate one latent space to another.