Hacker News Reader

Not so. They split the data between "bearded faces" and "non-bearded faces" and compute the vector from one to the other, and use that to alter a given face. (It reminds me of the typical word2vec example of man - woman + queen = king.)

See their code snippet halfway down their page, or "Semantic Manipulation" on page 8 of their linked paper.

View on HN

This sounds like yet another appearance of the tank-sky problem.

There is no mechanism in this paper (or in a standard VAE, or GAN) to encourage that a single human-understandable semantic quantity should be captured in a single dimension of the latent code. So, in general, it won't happen. "Blondeness" is spread out over all dimensions of the latent code. Therefore the method in this paper (summarized by matheist) is totally reasonable. There are other autoencoder-type generative models that try to concentrate each attribute in one dimension, usually by using the class labels as an additional input. But that is not the focus of this paper.

Do you have a link to a paper that does this by any chance? I'm interested in learning more but not sure what to search for.

Reply on news.ycombinator.com