Few-shot Image Generation via Cross-domain Correspondence
few-shot-gan-adaptation
Official repository for Few-shot Image Generation via Cross-domain Correspondence (CVPR ’21)
Utkarsh Ojha, Yijun Li, Jingwan Lu, Alexei A. Efros, Yong Jae Lee, Eli Shechtman, Richard Zhang
Adobe Research, UC Davis, UC Berkeley
Repository for downloading the datasets and generated images used for performing the evaluations shown in Tables 1 and 2.
Overview
Our method helps adapt the source GAN where one-to-one correspondence is preserved between the source Gs(z) and target Gt(z) images.
Sample images from a model
To generate images from a pre-trained GAN, run the following command:
CUDA_VISIBLE_DEVICES=0 python generate.py --ckpt_target model_name
Here, model_name follows the notation of source_target, e.g. ffhq_sketches. Use the --load_noise option to use the noise vectors used for some figures in the paper (Figures 1-4).