Depixelizing Pixel Art

[SynchronousArts]

Interesting technique for converting low resolution pixelated images:

http://research.microsoft.com/en-us/um/people/kopf/pixelart/supplementary/index.html

Direct Link to a PDF of the Paper:

http://research.microsoft.com/en-us/um/people/kopf/pixelart/paper/pixel.pdf

SA

[Astrophizz]

Nice, that actually looks like it's an improvement on hq4x in some cases.

[mike20021969]

The Ours results are amazing.
Is this "algorithm" going to be available to the public/available to purchase (maybe within a program)?

[LoRd_MuldeR]

The Ours results are amazing.
Is this "algorithm" going to be available to the public/available to purchase (maybe within a program)?

Well, they have a paper online (see first post in this thread), so in theory everybody could implement their algorithm in his/her own software.

However you'll probably have to buy a license to use their algorithm in a commercial application. Also it doesn't look like they provide a sample implementation...

[ramicio]

That would be awesome to implement in a NES or SNES emulator for today's HD resolutions.

[LoRd_MuldeR]

That would be awesome to implement in a NES or SNES emulator for today's HD resolutions.

If the algorithm is fast enough to run in real-time (on a normal CPU or GPU), then yes.

[Maccara]

If the algorithm is fast enough to run in real-time (on a normal CPU or GPU), then yes.

From the paper, 0.62s median per frame on a 2.4GHz single core. So no, not even remotely fast enough.

However, they also say code was not optimized.

[ramicio]

Wouldn't something like this benefit from running on a GPU rather than a CPU, once optimized? The info they give is not very detailed. Who knows what kind of CPU usage they were seeing (optimization), and single cores today is a joke.

[Maccara]

Dunno. haven't seen the code, but quickly browsing the paper I can see there might a few issues if you want to multithread this efficiently.

Even with perfect multithreading, a quad 2.4GHz would only yield 0.155s per frame. Still much too slow. And I wouldn't bet my money on feasibility of translating this to GPU processing either (looks bandwidth hungry to me).

If there was reference code, someone (not me) might be interested in looking at optimization possibilities. Until then this is moot speculation, unless someone wants to try to implement from the paper (but the algorithm details are a bit vague in that).

[Lyris]

Interesting, but for video games with pixel art, the nearest neighbor resizing still looks best to me.

[markanini]

The anti-aliasing is nice but hq4x looks more like the artist intended to me.