In traditional graphics work, vectorizing a bitmap image produces a bunch of points & lines that the computer then renders as pixels, producing something that approximates the original. Generally there’s a trade-off between editability (relatively few points, requiring a lot of visual simplification, but easy to see & manipulate) and fidelity (tons of points, high fidelity, but heavy & hard to edit).

Importing images into a generative adversarial network (GAN) works in a similar way: pixels are converted into vectors which are then re-rendered as pixels—and guess what, it’s a generally lossy process where fidelity & editability often conflict. When the importer tries to come up with a reasonable set of vectors that fit the entire face, it’s easy to end up with weird-looking results. Additionally, changing one attribute (e.g. eyebrows) may cause changes to others (e.g. hairline). I saw a case once where making someone look another direction caused them to grow a goatee (!).

My teammates’ FaceStudio effort proposes to address this problem by sidestepping the challenge of fitting the entire face, instead letting you broadly select a region and edit just that. Check it out: