That's a lot of info to go through but here we go. 

The iterative process 

Iterating over

Loaded into unity3d, adjusted lights, did direct downscaling - doesn't look good

Lanczos downscaling looked ugly

Area downscaling via python scripting

Adding an triangle depth normal edge detection + downscaling "lineart" pass, this one via lanczos because of lack of color detail, which made it decent

Almost painterly enough

Unwrapped, added green demontech plasma corrosion

Finally added 50% mixed soft posterization 

And finally everything together, with new sounds to boot.

Basic pipeline is this - I render at 640x400 in unity with sharpening, SSAO and posterization.

Then i render the plasma puff effect separately but without the alpha channel on a black background. Bring those over to a separate python script which creates alpha out of black channel, while also normalizing the color so it doesn't have colorbanding, resize it down.

Then I render the edges at 1280x800 separately, then bring it over to quite extensive python script which takes the sprite, scales it down, 50% multiplies it with the edges, then overlays it on top of the puff, after which it saves with the correct sprite name.

Sourcecode is included https://gist.github.com/smoketh/814bc095880bc48bc64058d52fb108c8

And that's pretty much it. 

A better method of core sprite (not outline) downscale has been also developed by me. I've trained a ESRGAN model specifically on subset of fakefaith 2x upscaled  vs nearest neighbor upscaled images.

Download link: https://mega.nz/file/Ex8HUJqJ#7k5FRc9qC-6CWSr6ih3TDy1WNl-RMCf8zbqSmZ5AFA0

I strongly recommend using CUpscale with this with NCNN core (it's tad better than cuda):

Pipeline is simple - throw sprites into cupscale and run 1x_doom_stable on it (you'll find setup instructions on github), and while at it - also select option to resize after scaling, pick nearest neighbor and 50%.