Depixelizing pixel art


bzar

A Commando
Joined
Sep 22, 2008
Messages
4,500
Location
Finland
Website
Visit site
Slashdot article


Mirror for the paper


Some of those results (like the dolphin in the first page) look unbelieveably good. I'm guessing this would take quite heavy processing. Maybe could be pre-processed? Ideas for applications? Emulator viability (maybe not on pandora, but in general)?


EDIT: Just to be clear, I have nothing to do with this paper :p
 
Last edited by a moderator:
Their results are good, I'd like to see some video to compare how it looks to other methods, as right now they only seem to be applying it directly to sprites, but many systems don't have sprites so I wonder how it looks when applied to a frame.
 
Seems to be very similar to how the hqnx family does it, but takes the algorithm one step further with its corner detection. In some cases that makes it an improvement, but it will definitely be quite CPU intensive. Possibly a candidate for offloading to the DSP, though I think for the two times scale we'd be needing, hq2x should be more than good enough.
 
As I understand, what makes this algorithm different from the hqnx family (and most other sprite upscalers) is that most upscale by a fixed proportion, while this new one instead vectorizes, allowing for any amount of upscaling or downscaling. So this technique is applicable when the resolution of the output device, and therefore the needed scaling factor, isn't known beforehand, or when the needed scaling factor is not an integer. The first condition is less important for Pandora's fixed resolution, but the second could be useful in some cases.
 
This one goes further in reducing of the biggest flaws I see in these types of upscalers, where a lot of disjointed curves are made when it was clear the source pixels weren't approximating that. But it still isn't perfect, so graphics like the C and Boo Diddly, that should look very round, still look a little lumpy. Even the dolphin shows some minor signs of this.
 
Back
Top