FEW-SHOT CROSS-DOMAIN IMAGE GENERATION

This work proposes a simple procedure to utilize a GAN trained on large-scale source-data to generate samples from a target domain with very few (1-10) examples. It is Capable of generating data from multiple target domains using a single source-GAN without the need for re-training or fine-tuning it. Given a GAN trained on a source dataset and a few samples from the target distribution, our optimizes a new multi-layer perceptron to output ‘novel’ target latent vectors. These are then fed to the source-generator to produce novel images from the target domain.

Mondal AK, Tiwary P, Singla P, Prathosh AP. Few-shot Cross-domain Image Generation via Inference-time Latent-code Learning. In The Eleventh International Conference on Learning Representations 2022 Sep 29