Um, I don't believe that would help. You either inline the images as data, or better, get rid of all images and have random length non-printing text to hide the real data. This kills caching and performance though.

https://github.com/technion/mod_randpad (an Nginx module that injects padding as a comment into returned HTML documents) is likely the correct layer to add random padding; the performance hit would be negligible.

SD dev here. You might be interested to check out https://github.com/freedomofpress/fingerprint-securedrop (note: still in experimental stages). We're going to be using the framework/ ML pipeline to evaluate server-side defenses (e.g., https://petsymposium.org/2017/papers/issue2/paper54-2017-2-s...), and are also working w/ Tor dev Mike Perry who is working on a solution or set of solutions to the problems of website & circuit fingerprinting with regards to Tor at the protocol level.

Edit: What Mike is trying to implement is based on https://arxiv.org/pdf/1512.00524.pdf.