Maxwell Render :: a new render engine

[jasonkey]

i think we could talk about this here, why not?

I for one, am very impressed with the gallery but a little pessimistic about the whole 'no artifact' thing. I've worked side by side with developers writing rendering engines and there is no such thing as 'artifact free'. Having said that, this does look like something more than marketing spin. Seems to be a much more physically correct approach.

I'm dying to try it.

[playmesumch00ns]

No artifacts? What about the bloody great noise everywhere in those architectural renders?

[Tommy5547]

Hmm I'm not convinced.. i think I'll continue with Brazil Render.

[tjnyc]

No marketing spin. It is an unbiased renderer, it is not new idea. The brief concept is that an unbiased renderer within a given amount of time will come to an error-free solution, artifacts are the result of errors in the solution which is an inherit problem with current baised renders. Most current renders interpolates, but cannot get rid of errors no matter how high you will have to bump up the settings for accuracy, you also have to rely on anti-aliasing to additional get rid of the artifacts. Even then there are NO gurantee that your result will be artifact-free. Maxwell purposes that within a set amount of time defined by the user, the render will try to compute the most accurate, error-free solution. So yes there will most likely be artifacts if say you set the final render for a minute, but by setting the time to any sufficient amount of time, it will produce an artifact-free result.



Cheers,

[Andrew W]

...but by setting the time to any sufficient amount of time, it will produce an artifact-free result.
Does sufficient amount of time equate to "nigh on infinite amount of time?"

A

[gga]

As tjnyc says, unbiased renderers are not new. Kajiya introduced the concept back in 1986, at the same time he presented the rendering equation.

So far unbiased renderers have remained more in the realm of theoretical rendering more than in the realm of being something useful, mainly due to their very long render times.
But it is also true that due to the complexity of their implementation they have not been explored that much. And it is true that the computers of 1998 are nothing like the computers of 2004.

Traditionally, unbiased algorithms are often used associated with montecarlo bidirectional path tracing with unlimited bounces or metropolis light transport.

If you want opinions on their usefulness...

You can, for example, find out both Peter Shirley's opinion and Eric Haines' in some old posts of ray tracing news:
http://jedi.ks.uiuc.edu/~johns/raytracer/rtn/rtnv10n2.html

For a more comprehensive analysis of these methods, you can read Eric Veach's thesis at:
http://graphics.stanford.edu/papers/veach_thesis/
where he presented MLT for the first time.

[cpnichols]

If you ever have used Lightscape, it was an unbias rendering engine. It would define the "error free" solution at 100%, and you would say what is "good enough," usually around 90%.

The progress woudl look somthing like this:
0%, 50% 75%, 82%, 85%, 87%, 88%, 89%, 89.5%, 90%

So technically speaking 100% is impossible.

[shehbahn]

perhaps more interesting is the fact that this renderer appears to be one of the few resolving its illumination in a spectral space instead of RGB. i would expect this to be on the very expensive side for a GI render...

[percydaman]

perhaps more interesting is the fact that this renderer appears to be one of the few resolving its illumination in a spectral space instead of RGB. i would expect this to be on the very expensive side for a GI render...
yeah but I bet itll make for beautiful renders!:thumbsup:

[playmesumch00ns]

Looking at it naively it seems to me it should be faster, no? I mean, your energy is represented as a single float, rather than a colour, so less memory and any energy transfer functions only have to operate on one value instead of three... of course, I'm probably missing something. :)

[gga]

A single float? Huh?

There isn't really a "right" representation for spectra, being infinite and all.
From the Poyton's Color FAQ...paraphrased....

Human perception considers the spectra between wavelengths of 380 to 780 nanometers.
Measurements of materials done with spectrophotometry result in a spectral power distribution (read: a table of measurements of radiance at intervals, frequently done at 5-10nm, through that visual range).
Thus, to represent that, you'd be looking at around 40-80 channels of data (Poynton mentions 31).

Obviously, that's not very feasible, so fewer wavelengths may need to be used.

Making things a tad worse, the human eye also has 4 different photoreceptors, each with a somewhat different spectral response curve. Fortunately, the CIE (Commission Internationale de L'Eclairage ) adopted some standard curves for your average person. These curves specify how a spectral power distribution can be transformed into a set of three numbers that represent a color.

So, at the very least, a spectral renderer would likely be using anywhere from 3 channels of data to 31 or more.

Not only that, but since you are not going to ask artists to suddenly start painting everything in a spectrum of X channels from now on, each texture fetch or rgb color you set in the gui of your 3d package will have to be transformed. And finally, your spectrum will have to be turned back into rgb for displaying the image.
For a discussion on the issues and algorithm to transform back and forth, refer to Brian Smits:
http://www.cs.utah.edu/~bes/papers/color/

[shehbahn]

you're reading my thoughts gga... are you involved with this kind of funky optics at all ?

[playmesumch00ns]

Aaah ok. I getchaI was just thinking about represting the colour. So basically each photon is a 2d function of wavelength and energy. So you've got to store and manipulate that 2d function, presumably as some sort of product of basis functions. Yep I could see how that would be very very slow. What we need is some sort of flux capacitor!

Thanks for the link gga. I'm just devoting some time to try and get my head round radiometry at the moment. Should spend some time reading up on optics too...

[shehbahn]

>So basically each photon is a 2d function of wavelength and energy.

pretty much yes - though the internal representation of the spectra is vague in their description.

[CGTalk Moderation]

This thread has been automatically closed as it remained inactive for 12 months. If you wish to continue the discussion, please create a new thread in the appropriate forum.