Procedural Noise using Sparse Gabor Convolution

[michael]

"In this paper we introduce a noise based on sparse convolution and the Gabor kernel that enables all of these properties. "

http://www.cs.kuleuven.be/%7Egraphics/publ...ns/LLDD09PNSGC/

[Jason]

"In this paper we introduce a noise based on sparse convolution and the Gabor kernel that enables all of these properties. "

http://www.cs.kuleuven.be/%7Egraphics/publ...ns/LLDD09PNSGC/

Nice find!

The only thing though - it's doesn't require texture coordinates.. so does it handle deforming geometry?

[Mario Marengo]

Very interesting!

Thanks for the link, Michael.

Yeah, the claim that it needs no surface parameterization sounds a little "too good" (the feature distribution looks like it's following gradients or something). But I guess the deforming "blobby cube" constitutes a deforming surface, so who knows.<shrug>

[sibarrick]

Nice find!

The only thing though - it's doesn't require texture coordinates.. so does it handle deforming geometry?

Just use rest coordinates for P and compute from those surely?

[Mario Marengo]

Just use rest coordinates for P and compute from those surely?

Except that would be 3D noise, not 2D. They claim it's a UV-free "surface noise [2D], complementary to solid noise [3D]", which I take to imply a parameterization-free 2D-like noise. But, I guess we'll find out when they post the actual paper...

[sibarrick]

well either way if you can calculate it on a static object if that object then starts deforming and you want the noise to stick you just store whatever parameters are needed from the static version, whether that is uv, P or something else doesn't really matter surely?

[sibarrick]

I see the code for this is now up, anyone tried implementing it in Houdini yet?

[Andy Nicholas]

Yep, you don't need UVs. As I understand it, It's doing a splatting type algorithm with small patches (they call them kernels) of noise with a user defined noise spectrum. It uses a local projection at each sampled location based on position and surface normal. You can see that fairly clearly in the video (the bit with the meta ball cube at 1:11)

Maybe a good analogy would be to think of using small squares of newspaper to do paper maché on a surface.

I think to get it to stick, you'd probably need an attribute associated with each sampled position on the mesh that doesn't change. It would use that as a seed to their random number generator which drives the noise and splatting. It could be a random number at each point or even the point number/ID, just as long as it doesn't change.

Anyway, that's how I read it from an initial scan of the paper. I might be wrong on some of the details.

Andy

[eetu]

Interesting paper.

I think to get it to stick, you'd probably need an attribute associated with each sampled position on the mesh that doesn't change.

I got the impression that one wouldn't necessarily want it to stick on a deforming surface.

As the surface deforms new noise smoothly fades in/out, always resulting in non-distorted texture.

Now that I write this, it sounds an awful lot like a solid 3d-texture

At least in the anisotropic case the difference should be clear, plus sounds like the filtering

is easier with surface noise than with solid.

I started thinking about implementing (means almost half done, right? the solid case, but realized

that I didn't know what the third dimension of the gabor kernel is supposed to be.. so it's waiting for a

trip to math-land.

eetu.

[Andy Nicholas]

I got the impression that one wouldn't necessarily want it to stick on a deforming surface.

I guess it depends what you're using it for. If I was using it in a skin texture for example, I'd sure as hell want it to stick

But I think it offers you the flexibility either way.

[eetu]

I guess it depends what you're using it for. If I was using it in a skin texture for example, I'd sure as hell want it to stick

But I think it offers you the flexibility either way.

Very true

I just happened to be thinking about the particle surfacer..

Anyhoo, hopefully someone gets to tackle this.

eetu.

[Andy Nicholas]

I just happened to be thinking about the particle surfacer..

Damn, yes. That's a good point. Getting textures to look as if they're sticking on fluids can be a real pain.

[sibarrick]

Yep, you don't need UVs. As I understand it, It's doing a splatting type algorithm with small patches (they call them kernels) of noise with a user defined noise spectrum. It uses a local projection at each sampled location based on position and surface normal. You can see that fairly clearly in the video (the bit with the meta ball cube at 1:11)

I agree, my reading of this is as follows:

They use a worley noise type cell algorithm so that for each shaded point you determine which cell the point is in and then you "splat" based on a projection of 2d noise centered around that cell using the shaded point normal to determine the 2d plane for the splat projection, you then do the same thing for the each of the neighbour cells and combine the results weighted by the distance the point is from the center of the cell in question.

For isotropic noise they orient the 2d splat planes randomly for each grid cell and for anisotropic they use a flow direction to orient them coherently.

I think this part of the paper is very interesting and could be separated out from the noise generator to be used with any existing noise function and I might have a crack at that....

Their noise function is also very interesting but I not sure yet how easy it would be to control without their fancy interface, since the final results of it seems on the face of it to be very similar to existing noise patterns I'm wondering whether the only advantage of it is that one function fits all, but that is negated by the complexity of the parameter interface required to control it. Therefore the thing to start with here is the splatting algorithm.

[kleer001]

Whoa, am I late to the party or what?

News here actually. There's a new paper expanding the Gabor Noise hooha (published about a year after the last post here)

(down at improved Gabor Noise)

http://people.cs.kuleuven.be/~ares.lagae/