Digitally-upscaled remakes of classic pixel art games are usually a crime against eyeballs, but these incredible Doom sprites—generated with the help of neural networks—are proof that maybe that won’t be the case forever.
Hidfan over at the Doomworld forums has run original Doom textures through two different neural networks specifically designed for image upscaling: letsenhance.io and Nividia’s GameWorks SuperResolution.
The textures are artificially upscaled to 8x, then downscaled to 4x, then downscaled again to 2x fit the pixellated aesthetic, with any rough edges cleaned up manually.
The end result is a game that, while maintaining the spirit of the original game’s aesthetic, is now looking a lot sharper:
If you want to try it out yourself, this thread will get you started.
DISCUSSION
This is because Doom wasn’t “pixel art”. They were low-res, like pixel art, but they were digitized images - photos of some clay sculptures that had been made. There was some editing done later (all the possessed were made by editing the Doomguy images, a lot were recolored as well) but there’s an underlying real-world source for them.
That’s why they don’t actually look bad with simple bilinear upscaling. And it’s much easier to train an AI to interpolate details that are essentially photographic.
Pixel art, in contrast, has hard lines defined on pixel boundaries. Upscaling algorithms need to figure out where they can blend between two pixels, and when they can’t, and when they get it wrong, it looks awful. The only reasonably good-looking pixel art upscale, besides nearest-neighbor, are ones that don’t do any blending (no pixels have a color that was not in the original), but instead just change the shape of each “square” to flow better.