Distance field generation

Adobe abandoned Pixel Bender so I looked around for
a better platform for distance field generation. Filter
Forge was an obvious choice.

Has anyone attempted to implement it here? The Pixel
Bender version ran on the GPU so having each pixel check
neighbors in circles of ever increasing radius was a
workable solution. However FF is a very different beast
so I'm not sure if there are some theoretical limits
preventing a proper algo from finishing in a reasonable
time.

Thanks although I'm not sure what that is.

Here are a few calculations:

Let's assume a 512x512 image with max distance of 256 and wrapping around at the edges.

We have 262144 pixels, each pixel will check less than 256*256 = 65536 neighboring pixels. That would be checking a square but we only need a circle around the target pixel. The area of a circle of 0.5 radius is 0.78 so that's gives us 65536*0.78 = 51000 pixels with a radial search.

So the worst case scenario is that logic runs 262144*51000 = 13369344000 times and we want it to finish reasonably quickly, let's say in 5 minutes or 300 seconds. That would be 44564480 sample evaluations a second. 44.5 million... that sounds a lot... even if it will almost always be less than that.

Good finds! I'll mock up some dummy sampler to see how many raw samples I can get out of this hardware each second.

-snip-

Wrong thread, sorry ;D

Ah good to know... I still haven't had time to work on it. :\

Also, I don't know if the tiny blur deteriorates the output critically. I guess that depends on the application. Since the blur is in the same direction as the ramp, I doubt it makes any real difference in most cases.

Bevel at 256 pixels, around 24 minutes for a surface filter.

I took a stab at it yesterday:



Now I'm trying to debug why the far away pixels suddenly become white...
The sampling is done on a square area at this point.

Alright, I got this working. It's a 600-800 times slower than the PixelBender version but it doesn't crash the video driver at high iterations which is good.

On a 1K image 256 pixel search radius takes hours to render but 64 is just a few minutes. (The radial search algo still has issues with standalone single pixels but I'll fix that later.



Using a bitmap component before the script node (blur with (near) 0 radius) made the letters grow some really weird appendages... not sure what's going on there.

I'll do a quick optimization pass then submit it.

Yes, that's my blog. PixelBender was abandoned so it doesn't run in newer Photoshops. Plus it could kill the video driver so often you had to calculate stuff one area a time, so the workflow was rather crappy.

I did try minimum/maximum some time ago but it didn't produce a usable result (ugly distortions as I got further away from the original contours) and it was still extremely slow.

Ok, here is the filter in case anyone wants to take a stab at speeding it up:

DistanceField.ffxml

Might work in certain cases but the distortions introduced mess up text related effects for example.

It seems that the algo is inherently unfit for FForge... I'll try Processing next.

New version with calculation restricted to a 256x256 cell. Cell position can be adjusted so one can process an image in several passes.

DistanceField.ffxml