Upscaling Retro 8-Bit Pixel Art To Vector Graphics 325
An anonymous reader writes "Two researchers — Johannes Kopf from Microsoft, and Dani Lischinski from The Hebrew University — have successfully created an algorithm that depixelizes and upscales low-resolution 8-bit 'pixel art' into lush vector graphics. The algorithm identifies pixel-level details (original paper — PDF) to accurately shade the new image — but more importantly, the algorithm can create smooth, curved contour lines from only-connected-on-the-diagonal single pixels. At long last, we might be able to play Super Mario Bros. on a big screen without stretching our beloved plumber's pixels to breaking point. You really must look at the sample images." Scroll down in the paper to see how their technique stacks up against some others, including Adobe's Live Trace.
Re:PDF slashdotted (Score:4, Informative)
http://johanneskopf.de.nyud.net/publications/pixelart/paper/pixel.pdf
Don't tell me people have never heard of the coral cache
Paper mirror (Score:5, Informative)
My verdict: Yeah, it looks all nice and smooth, but with all upscaling, it's basically interpolating data. The original just didn't have that much detail, so you can only get so much out of it. Sometimes it works, sometimes it doesn't.
(Oh, and it makes all text look pretty bad. Kinda Comic Sans-y, if I can say that without invoking instant hate.)
Re:Alright... (Score:4, Informative)
Re:This is news? Anyone else run a NES emulator? (Score:5, Informative)
RTFP - they list the various existing upscaling algorithms, and why they're not perfect:
Several later algorithms are based on the same idea, but use more sophisticated logic to determine the colors of the 2x2 block. The best known ones are Eagle [Unknown 1997], 2xSaI [Ohannessian 1999], and Scale2x [Mazzoleni 2001], which use larger causal neighborhoods and blend colors. Several slightly different implementations exist under different names, such as SuperEagle and Super2xSaI. An inherent limitation of all these algorithms is that they only allow upscaling by a factor of two. Larger magnication canbe achieved by applying the algorithm multiple times, each time doubling the resolution. This strategy, however, signicantly reduces quality at larger upscaling factors, because the methods assume non-antialiased input, while producing antialiased output.
The latest and most sophisticated evolution of this type of algorithms is the hqx family [Stepin 2003]. This algorithm examines 3x3 pixel blocks at a time and compares the center pixel to its 8 neighbors. Each neighbor is classied as being either similar or dissimilar in color, which leads to 256 possible combinations. The algorithm uses a lookup table to apply a custom interpolation scheme for each combination. This enables it to produce various shapes, such as sharp corners, etc. The quality of the results is high. However, due to its strictly local nature, the algorithm cannot resolve certain ambiguous patterns and is still prone to produce staircasing artifacts. Lookup tables exist only for 2, 3, and 4 magnication factors.
whereas what they did here is:
Our goal in this work is to convert pixel art images to a resolution-independent vector representation, where regions with smoothly
varying shading are crisply separated by piecewise-smooth contour curves.
Seriously, just look at the whale image linked from TFS.
Re:Yawn (Score:4, Informative)
Comment removed (Score:5, Informative)
Re:PDF slashdotted (Score:5, Informative)
CORAL link to the PDF (Score:5, Informative)
It looked like it was on the verge of a Slashdotting, so I CORALized it.
http://johanneskopf.de.nyud.net/publications/pixelart/paper/pixel.pdf [nyud.net]
Very interesting, and quite effective.
Re:Paper mirror (Score:5, Informative)
Re:Great. (Score:4, Informative)
The maximum theoretical human eye resolution (limited by aperture and focal length) is about 0.35mm per line pair at 1m. For most people with "normal" eyesight, the limit is closer to 0.7 mm at 1m. So for a TV that's 6 feet away and 2 feet tall, there's no point to a vertical resolution higher than ~1800 pixels, and 1080 is more than what most people can perceive. 4320 pixels at 6 feet makes no sense unless your TV is a minimum of 5 feet tall, and for most people, 15 feet ;) So yes, it really does take wall-sized TVs for that sort of resolution to even be useful. A 7" panel (~4" tall) with 1600 pixels vertical resolution? That's 0.06mm per pixel. The maximum theoretical distance in which that could be useful is under 7 inches, and more practically, closer to 3 inches away from the viewer. If you want that kind of resolution, you might as well make it into a planetarium ;)
Re:Wonderful! (Score:5, Informative)
It is. This has nothing to do with that, but is about giving a nicer representation to information that is there. All this does is removes jagged corners that result from up-scaling.
Sorry, but that's just plain not going to work.
Re:Sample image is not 8bit source (Score:4, Informative)
Yeah, it's from Super Mario World, i.e. the SNES - and on the SNES [wikipedia.org] "Sprites can be 8 × 8, 16 × 16, 32 × 32, or 64 × 64 pixels, each using one of eight 16-color palettes", which would make the dolphin a 4-bit sprite.
Re:Wonderful! (Score:3, Informative)
From TFA : "The other problem is that the Depixelizing Pixel Art approach always smooths images, even when an object shouldn't necessarily be smooth. For example, are Space Invaders really meant to be cute and round? Maybe, in the creator's eye, they had long, angular, razor-sharp mandibles and straight-out antennae! "
They are adding information, but they have no way to know if it is what the original artist had in mind.