I'm writing my dissertation (music theory) with LaTeX; the text is all stored in git, and all of my examples are PDF files in a submodule (so the main repo doesn't get huge). I need to have the examples tracked, and the submodule allows me to pin a particular version of the examples with a particular version of the text. If I need the PDF of a chapter I sent to my advisor, for example, I can just check out that tag and run xelatex again to generate the real document.

The example PDF files aren't easily generated with TeX, so I make them in Illustrator and store them with Git. They change a little less frequently than the text itself, and usually once an example is done it doesn't get updated very frequently. In my case, at least, I can't easily version the "raw source" for the examples, as the Illustrator binary files are ~3x larger than the resulting PDFs.

I suppose I could use something like git-annex, but every time I've tried I haven't really been able to get the hang of it, and the time I'd spend learning it is better spent actually writing the thing.

(Incidentally, this new Github feature makes it much easier to browse my examples repo!)

Versioning binary files is a wasteful, but people still frequently need to do it (or at least put a static, unchanging PDF into a repository). For example, I might get set of feature requirements in PDF form and include that in my repository just so anyone who clones it can have the same requirements file that I do.

In cases where you have to include a PDF, it's very helpful to have it be viewable inline on a page.

The alternative is to store pdf-from-client-v1.pdf, pdf-from-client-v2.pdf, etc

That doesn't address the proposed alternative: version the actual source file (eg: lated?) rather than the PDF file.

I was really hoping for them to implement this, because I like to keep relevant scientific papers next to the code. If for example I were to implement an algorithm described in a publicly available paper, I'd much rather have it sit right there in the repo instead of having a link to some other site (which may or may not be reachable at any future point in time).

Also I view PDFs as only "half binary". It's just a bunch of text streams, some of which may be compressed with deflate. (Granted, there's also stuff like embedded images, fonts, etc.)

Design comps/wireframes are frequently shared as pdfs. Designers do not like storing their source files in git.