New algorithm promises to make retro-pixel games smooth again?

Kaotik

Drunk Member
Legend
Supporter
http://johanneskopf.de/publications/pixelart/index.html

The link isn't functional at the moment, but these pics were from there:
X-20110526152945453.jpg


X-20110526152949750.jpg


Pretty fancy looking, if indeed true - but since the link is down I can only trust the word of the media reporting this.

According to Dome.fi, researchers Johannes Kopf and Dani Lischinski are behind this new algorithm, which could bring new life to many retro games
 
Haven't had a chance to read the paper yet, but looked over some of the image comparisons they have online and it looks quite good.

It does have a problem with more complex shapes, like Mario and Yoshi. But that could possible be alleviated combining this technique with a prepass with another technique, like bicubic or hq4x, that gives better/easier edges to work with.

You can find the full paper here:
http://research.microsoft.com/en-us/um/people/kopf/pixelart/paper/pixel.pdf
And image comparisons here:
http://research.microsoft.com/en-us/um/people/kopf/pixelart/supplementary/index.html
 
It definitely needs an intermediate step though as it tends to destroy a lot of info on some things the hq4x retains.
 
That's pretty awesome! Works extremely well with the line-art cartoon of many games, and gives a very soft, organic feel to the graphics. They could do with guessing gradient fills. Take the Yoshi example - the pixels are a green gradient, but the result is green banding. Given a suitable similarity, they could interpolate shades of the same colour to better effect. It also struggles with soft detail in the Golden Axe character, which comes out the worse for this scaling. All in all though, very impressive for making old games playable on modern equipment - certainly 'cutesy' games.
 
Would the new algorithm applied on hq4x'ed image do better? Would it be too heavy?
 
Maybe. It depends on if hq4x is better at finding patterns.
 
Last edited by a moderator:
theres a big difference to hq4x and other upscaling algorithms. upscalers just look on a couple neighboring pixels and can be run on parts of the screen as well.

this one converts the whole image to vector graphics, means every pixel could affect a big area. Not really suited for upscaling moving pictures, its more for converting sprites into vector graphics then clean them up a bit by hand and use them directly in a port.
Not to mention its really slow
 
Back
Top