4 ms·

Probably as a research project to see if it's worth using in HD conversions of old games, such as the Age of Empires remake that's dropping tomorrow. The algori

by jjcm 9y ago

Probably as a research project to see if it's worth using in HD conversions of old games, such as the Age of Empires remake that's dropping tomorrow. The algorithm seems to work well on standalone sprites, but kinda merges in a gooey way when two sprites are next to each other. The super mario example from the previous discussion shows this slightly: http://files.jjcm.org/upscaling.mp4 http://files.jjcm.org/upscaling.mp4 Very curious to see this as it progresses however.

ttoinou 9y ago
Taking time into account is missing. I guess the graph reconstruction could also depends on some kind of pixel art optical flow
simcop2387 9y ago
I imagine it was also useful to see if it could be used to scale up low res artwork in older programs to hidpi displays.
pradn 9y ago
It looks like the transformation in being applied to the whole image at once in your example. If each sprite was vectorized individually, you wouldn't see the gooeyness. If you are a game developer, this change should be easy enough.
simooooo 9y ago
I'm certain I've already seen this as a shader option in emulators
c12 9y ago
I'm still impressed with the super mario example, I can't personally see anything negative about it and would enjoy playing the game with that "art style." It's not perfect but seeing the alternatives it is a huge step in the right direction.
LoSboccacc 9y ago
I don't think it was about videogames, more likely upscaling algorithm would have saved a TON of bandwidth for videos. a a matter of fact some of the most recents algo available for video deinterlacing like nnedi3 are perfect for that, and of course also work for pixel art, and I think they blow away the example showed in their gallery - https://www.youtube.com/watch?v=0691zsXWbhA https://www.youtube.com/watch?v=0691zsXWbhA I wonder if now, in the tensorflow era, a better automatic approach using generative nn could be devised to reach beyond 8x

by jjcm 9y ago

ttoinou 9y ago

Taking time into account is missing. I guess the graph reconstruction could also depends on some kind of pixel art optical flow

simcop2387 9y ago

I imagine it was also useful to see if it could be used to scale up low res artwork in older programs to hidpi displays.

pradn 9y ago

It looks like the transformation in being applied to the whole image at once in your example. If each sprite was vectorized individually, you wouldn't see the gooeyness. If you are a game developer, this change should be easy enough.

simooooo 9y ago

I'm certain I've already seen this as a shader option in emulators

c12 9y ago

I'm still impressed with the super mario example, I can't personally see anything negative about it and would enjoy playing the game with that "art style." It's not perfect but seeing the alternatives it is a huge step in the right direction.

LoSboccacc 9y ago

I don't think it was about videogames, more likely upscaling algorithm would have saved a TON of bandwidth for videos. a a matter of fact some of the most recents algo available for video deinterlacing like nnedi3 are perfect for that, and of course also work for pixel art, and I think they blow away the example showed in their gallery - https://www.youtube.com/watch?v=0691zsXWbhA https://www.youtube.com/watch?v=0691zsXWbhA I wonder if now, in the tensorflow era, a better automatic approach using generative nn could be devised to reach beyond 8x