Fascinated by the concept! Machine-learning-anything always grabs my interest. Not sure I can offer much help (and it might be too late anyway) but I remembered seeing this guy's shader for Blender that converts 3D models to pixel art:
http://www.cgchannel.com/2015/02/this-free-blender-shader-turns-3d-models-into-pixel-art/Sure it's not exactly "photographs" but it would provide some kind of streamlined method of feeding in full-res textured models and rendering images with the pixelation shader toggled on and off to create the before-and-after photos you'd need.
There's another guy who seems to have mastered the hard lines of pixel art being rendered live in his game engine here:
https://www.youtube.com/watch?v=h5SjwrJen3o That sort of defined line work around objects is one of the tougher things to simulate with automation (I'd imagine), but again using the same method above (assuming you had access to whatever engine this developer is using, and how he's pulling it off) you could render images from 3D modeled scenes in full res, and then again with the pixelation shader. Might be just enough to guide the machine learning along, but hey what do I know!
Good luck and be sure to post again if you do manage to come up with something!
