Path Space Regularization (aka the solution to SDS paths)

[lacilaci]

If I understand this right, the disappearing caustics happen cause when scale becomes too low then metropolis cannot find paths? Or in your cases caustics never disappear?

And it seems clamping works ok I think... I had autoexposure enabled and probably the sun reflection was taking over

[lacilaci]

Could it be that the problem is related to bump mapping? In Dade's pool, the mesh surface is displaced, while lacilaci's water surface seems to make use of bump mapping.

Even on a reflection on flat surface I had caustics slowly disappearing on me...

Also, the smaller scale gets the less it is going to appear on rounded surfaces, so you cannot use large scale cause you would have very blurry caustics, smaller values will make caustics not appear on rounded surfaces and very small values that could produce super sharp caustics are very noisy...

[Dade]

This thing is starting to work (denoised lower right corner):

There are 2 important factors at the moment:

#1 you have to trust the algorithm, it is important to start with blurred caustics so Metropolis can find all of them and the algorithm will than sharpen them. Compare the above rendering with the initial blur:

See ? After a while the caustics are a LOT sharper.

#2 if you render for "too long", the angles will become too small (i.e. too much "sharpening") and samples will start to be all black and the image will start to be darker and darker.

I'm quite ok with #1 but #2 is a pain and must be solved. Some kind of lower cap could do the trick.

[Dade]

If I understand this right, the disappearing caustics happen cause when scale becomes too low then metropolis cannot find paths? Or in your cases caustics never disappear?

Yes, see the above post.

[marcatore]

#1 you have to trust the algorithm, it is important to start with blurred caustics so Metropolis can find all of them and the algorithm will than sharpen them. Compare the above rendering with the initial blur:


psr2.jpg

See ? After a while the caustics are a LOT sharper.

#2 if you render for "too long", the angles will become too small (i.e. too much "sharpening") and samples will start to be all black and the image will start to be darker and darker.

I'm quite ok with #1 but #2 is a pain and must be solved. Some kind of lower cap could do the trick.

I can't understand which is the image with more samples... the foreground image seems the longest rendered cause it has less noise but speaking about sharpening, it seems more or less the same... or even more blurred.

About the #2 , as cap you mean to limit how much small will be the search angle?

[lacilaci]

This thing is starting to work (denoised lower right corner):


psr1.jpg


There are 2 important factors at the moment:

#1 you have to trust the algorithm, it is important to start with blurred caustics so Metropolis can find all of them and the algorithm will than sharpen them. Compare the above rendering with the initial blur:


psr2.jpg

See ? After a while the caustics are a LOT sharper.

#2 if you render for "too long", the angles will become too small (i.e. too much "sharpening") and samples will start to be all black and the image will start to be darker and darker.

I'm quite ok with #1 but #2 is a pain and must be solved. Some kind of lower cap could do the trick.

Well.. my question is, if you cap it what would be the limit of the sharpness?

Cause I could get pixel wide caustic "lines" with PSR underwater (but too noisy and under some angle not even visible) but I worry that this is not very doable (in a reasonable time) even with the "cap"

And I cannot help myself, but that darkening issue - don't you think that it kinda screws up with the results overall as well?

[Dade]

I can't understand which is the image with more samples... the foreground image seems the longest rendered cause it has less noise but speaking about sharpening, it seems more or less the same... or even more blurred.

The 2 windows inside the single image are raw Vs denoised. You have to compare the first posted image with the second.

About the #2 , as cap you mean to limit how much small will be the search angle?

Yes.

[marcatore]

The 2 windows inside the single image are raw Vs denoised. You have to compare the first posted image with the second.

Ok. Now it's more clear.

Really not bad.

[Sharlybg]

See ? After a while the caustics are a LOT sharper.

How fast it is ?

[lacilaci]

So the smaller is the scale the more it is important that the polygons directly face the camera. The moment there is some angle between camera and polygons it will take forever to find some caustics there.