Maxwell Render | Official public forum

I know noise in maxwell render is a big problem, but because of its approach I also think it can't be totally avoided, though it can be reduced; tools like Neat Image works by selecting visible noise parts in the image to build a profile based on the current image... after that, this modeled noise is deconvolved from the image, resulting in far better results than standard generic noise removal tools like the Kodak plugins or the Photoshop builtin filters; on the Neat Image website there are many downloadable noise profiles taken from the most popular digital cameras (they are obtained taking a shot of reference image)

from this point of view, one nice thing to do would be to make maxwell render an image similar to the reference one (http://www.neatimage.com/testtarget.html) and build a "Maxwell Render profile" just like any other cameras; this would be useful, because not every image contains enough flat areas with noticeable noise for the program to learn. in other words, this could be intended as the "Maxwell noise model", based on the effect of rendering as many surface types as possible...

just this one should be a great step, but the further one could be to integrate into Maxwell this kind of deconvolution based on its peculiar noise; who better than developers themselves can know the internals of their engine to give a precise "noise pattern"? like in Lightware option, there could be in Maxwell the possibility to apply this custom deconvolution filter based on this precise profile, without the need to make post noise-work in Photoshop.

I don't know if it's possible to be implemented in the current maxwell version, but it should be a nice idea for the future... someone agrees with that?

... and the great advantage of a build in denoiser could be, Maxwell know, what an edge is and what not if it look in the scene geometry. A denoiser don't know, is it noise or a small edge.

The problem of a noise profile could be, that Maxwell produce different kind of noise at different sampling levels. Or?

this could be easy solved, because, Maxwell profiles could be made for various sampling levels, exactly like cameras profiles exist for various ISO levels.

tomorrow sure I'll make Rhino models to emulate the reference image for various sampling levels to build a preliminary profiles and I'll post some images...

DELETED

yes, I also think the internal denoiser would be a waste of time for NL programmers now, because it's better they concentrate on release version; in fact, by now I thought it would be a nice experiment make Neat Image learn Maxwell's noise giving it as much examples as possible, refining it for each one (many colors, glass refraction, caustics, shadow outlines, etc...)