Right. What it really does is take an image of a character in an arbitrary, but mostly standing, position, and returns an image in a standard resting stand position. This is neat, but costume generation is several steps past that.

Now, if you could take a collection of pictures of a character from anime and get a 3D model you could feed a fashion CAD system, that would be impressive.

There is CAD for fashion, and it's very good.[1] The CAD systems understand cloth physics quite well now. A machine learning system which takes a set of pictures of a character in different poses and tries to make clothing which matches would be a useful achievement. It's a big job, but probably do-able. The problem is reasonably convex. If you have a character in a dress, and you start with a generic dress in the CAD system, you can keep tweaking parameters and textures until it matches. The ML system doesn't have to invent a 3D dress, just keep tweaking a basic form.

[1] https://clo3d.com/

Same, possibly by taking 3d models and flattening them or something. Paper is interesting none the less.