As a deep learning-based image generation method, GAN produces a lot of amazing results. In particular, by changing the latent vector after learning, it is possible to make changes that have a number of meaningful meanings, so it makes you think that the latent space is not simply a random vector space but expresses the visual meaning of the image well.
However, it is not easy to analyze how changing the latent vector in the latent space brings about an interpretable change, and conversely, how to change the latent vector in order to make an interpretable change.
In GANSpace, a joint research project of Aalto University, Adobe Research, and NVIDIA, as an approach to solving this problem, PCA is applied to the latent space or similar modified vector space to extract a plurality of principal component vectors, and then the latent vector in the direction of the principal component vector It has been shown that you can easily create a visual transformation in the desired direction by changing the. Research was conducted on StyleGAN and BigGAN, and slightly different techniques were applied due to the structural difference between the two models, but basically similar methods of control were possible. In particular, since PCA is applied to the latent space or feature space of the previously trained model, there is no need to retrain.
In the GANSpace github repository, interactive demos and tools to visualize analysis results are publicly available, and GAN models support ProGAN, BigGAN-512, BigGAN-256, BigGAN-128, StyleGAN, and StyleGAN2. Here is a link to the GANSpace github repository.