Official Off-Topic Thread 2014

[Carnivac]

Great work, Carnifax.

 

[Vakinox]

Great work, Carnifax.

 

Maybe he thinks your name is a reference to the metal band: Carnifex?

Or the Latin word for "executioner"?

Or I could just be overexamining a simple error :p

[Seiseki]

Great work, Carnifax.

 

"oh no he didn't..."

(I love these smiles )

[Cyangmou]

http://www.messybeast.com/freak-face.htm

know ya kitten

[Ai]

A project to implement the Kopf-Lischinski pixel art vectorization algorithm in Inkscape was recently accepted for this year's Google Summer of Code.

The Kopf-Lischinski research was posted here before, but if you don't remember, this link gives a big list of examples of how the algorithm performs on familiar sprites. 
And this sprite might also remind you:



To me, the results of this algorithm seem clearly the best (including against Inkscape's currently built-in vectorizer "PoTrace" which performs excellently on everything but pixel art); my only concern is the addition of blur when colors are detected to be sufficiently similar. Hopefully that is a threshold that could be exposed in Inkscape for tweaking.

While I find upscaled pixel art to have a 'coarse' feel (lack of fine details feels odd at increased scale), this is certainly my best bet for doing auto-AAing as a postprocessing step (ie. vectorize with this new algorithm, then render at the original resolution)

[ptoing]

Still not a fan of how any of this looks at all, even though this one is better than PoTrace for example.
To me all of this looks like more or less successful attempts to convert stuff into window-colours. And everything gets round and blobby, no matter what. I'd rather have huge fucking pixels with some scanline filter than this, any day.

[Ai]

All upscaling that is anything but nearest neighbour looks wrong to me. But who cares about upscaling, I just want effective auto-AA as a post-processing filter. That's what this is *really* good for.

[YellowLime]

All upscaling that is anything but nearest neighbour looks wrong to me.

But who cares about upscaling, I just want effective auto-AA

Touché, Seiseki (jk)

[Dusty]

I'd rather have huge fucking pixels with some scanline filter than this, any day.

As a pixel artist seeing games rendered in their full nearest-neighbor pixels doesn't bother me. In fact that's how I prefer seeing pixelled games... but a lot of people, even old gamers, apparently can't stand the sight of pixels. I was surprised how many of my friends use scaling filters on their emulators.

I would like to see that scaling algorithm applied to a screenshot of a game as a whole, rather than individual sprites. Chances are since it was developed by Microsoft it was mainly intended to be used to upscale the old arcade games they port to their Xbox Arcade.

You can see on this sprite that has way more pixels happening:



The algorithm kind of craps itself.



I can only imagine what an entire game would look like, if it rendered out like that.

[Ai]

Dusty:
No need to imagine! Just look at the Super Mario World clip on their website -- In the 'Video Comparisons' section, click on 'Our Result'. The clip is limited in area, but certainly has enough to get a decent idea.

Overall this would confirm Ptoings statements, except the ones where he's merely being grumpy . It does look overly rounded to me (and there's an amusing 'slime blobbing together' effect as Mario + Yoshi's outlines animate as he rides Yoshi.) but it doesn't seem to suffer from the intentional-blurring found on the static sprites -- maybe they turned it off because it was too CPU intensive to render. Some people might like the result, but anyway it's moot for the immediate future since it's beyond the capabilities of current hardware to render in realtime.

Also: it actually performs very well (meaning better than all other algorithms -- the website allows you to easily compare.).  No algorithm is ever going to be able to 'guess' the intent of an isolated pixel that well (barring use of neural nets and other seriously heavyweight stuff), so it's much more accurate to judge on the basis of the ratio of pixel clusters that it succeeds/fails to detect.

I'll reserve my judgement on the blurred areas, as I can't tell how well/badly they have been vectorized.

Overall I'd like a hybrid version that also does the kind of layered tracing that Inkscape implements (treat the image as a stack of colored paper cutouts ranging from darkest-color to lightest-color. Note that this is NOT the approach shown in the comparison with PoTrace). This approach can result in some discoloration of edges but is somewhat better at conveying volume (since it essentially traces masses of light). Edge discoloration would be insignificant if rendered using no AA at a high resolution then downsampled to the original resolution.

Well, to get a really satisfying result would probably require two passes, with manual adjustment inbetween (vectorize -> patch, render as hires bitmap -> vectorize). If Inkscape was faster on complex images, it would be no big deal to draw a handful of shapes to clarify the vectorization, over the top of the pass1 result. Even after that there's the 'missing small scale elements' factor that's unavoidable. Even if I can get a perfectly accurate upscaling (definitely possible with simpler shapes) things just feel coarse and it's disappointing.