That's certainly a nice problem to work on. The major problem is, in most cases, getting enough paired data (human drawn pixel art for 3D renders) or a sizeable amount of similar-style pixel-art and 3D renderings for some sort of unpaired translation (e.g., CycleGAN).

Less time-consuming would be to take a look at style-transfer and see if such networks could pixelate a 3D rendering to some usable extent, but I wouldn't bet on this yielding good looking pixel art.

If you actually have a use case / dataset and you would like some insight, feel free to call me on LinkedIn.

Thanks for reading o/