4 - Final gathering with Global illumination.

 

How to achieve a clean rendering when Final gathering "needs" Global illumination?

 

As I mentioned above, the Final gathering method used alone can achieve really clean results. That is, almost perfect if you can stand huge rendering times. What I was complaining about before we discovered this way of using FG, is that Mental Ray's Final Gathering used in conjunction with Global illumination always lead to imperfect images. That is, as I already described, a generally satisfying result of lighting behavior (physically correct mimic of reality) but with here and there inconsistent lighting at corners and shadow artifacts that come up every now and then.
Since we better understand now how Final Gathering really works and what kind of result we can expect from it, the next step would be to get the same quality when Final gathering "needs" Global illumination.

 

Here are the statements :

 

Final gathering method used alone :
- When you put an object inside a constant sphere, as far as I understood, Final gathering casts rays in all directions for each pixel of the object. The more a pixel "sees" the sphere (which acts like a light, in this case), the more Final gathering gathers energy from the sphere and the more the pixel is lit ; on the contrary, if a pixel is situated where objects hide a lot of the rays that can be casted toward the sphere, this pixel will only be partially lit, etc...
- To make it simple, we can say that in these circumstances (diffuse lighting), the energy doesn't travel across the space. Objects don't transmit any light, they only stop the diffuse lighting, bringing smooth shadowing under them.
- The "key" parameter is the Max rad. I observed that the noise seems to have a scale. That is, when you set the Max Rad to 0.1, you tell the Final gathering to take into account any geometry that fits into 0.1 (unit of Softimage). In other words, if you have a cube of 0.1 by 0.1 among other bigger objects, and you set your Max rad too far from 0.1 (lets say 0.5 or more), you will inevitably get some imperfections around the little cube. As Luc-Eric Rousseau explained, the accuracy only blurs these imperfections.

 

Final gathering method used in conjunction with Global illumination :
- When you're not in the case where objects are lit by a diffuse sky but rather by one or several lights, that is in pure interior situations, the light doesn't come constantly and from everywhere. The light comes from one point (or several), crosses the space and bounces over objects until all the energy is absorbed. This phenomenon is computed from Global illumination, that is, giving the light an energy and giving objects the ability to receive, absorb and transmit this energy. All this relevant information is collected into a Map (the photon Map) that Final gathering invokes to do its job the same way it does without Global illumination.
- Since we know how to handle with Final gathering, that is we understand the way it works, we need to find a reliable manner to define Global illumination's parameters in such a way that Final gathering could bring the same clean results. That's another story !

A typical night and day Global illumination scene :

 

In the 2 images below, I tried again to better the rendering of a scene that needs Global illumination, starting from this better understanding of the Final gathering part of the problem.
On the left hand image, I use a simple architectural space lit by the sun (a distant spotlight) entering through all kind of windows.
On the right hand image, the same space at night, only lit by one interior spotlight.
In both scenes, I have a constant sphere outside that acts like a sky (day and night).
All the objects receive and transmit energy. Until now, I never really saw a difference of rendering time when only setting some of the objects as transmitters and receivers (let's say only the useful ones, that is ceiling, floor and walls to simplify). Perhaps having all the objects set as transmitters and receivers is not a general good rule to get fast renderings, but it's always hard to decide what object is going to be considered as negligible and generally, there are few. In these 2 scenes, only the windows' frames could be.

 

 

Here are the parameters for the left hand image (daylight) :
Global illumination : Spotlight intensity = 50,000 / Exponent = 2 / 500 photons / Photon depth = 6, 6, 6 / Accuracy = 200 / Radius = 4
Final gathering : Accuracy = 800 / Max rad = 0.04 / Min rad = 0.02
Rendering : size = 1280 x 984 / time = 3 h 25 mins / 4 threads / computer = dual PIII 733 (256 Mb Ram) + dual PII xeon 450 (512 Mb Ram)

 

Here are the parameters for the right hand image (night) :
Global illumination : Spotlight intensity = 300 / Exponent = 1.6 / 2000 photons / Photon depth = 8, 8, 8 / Accuracy = 200 / Radius = 4
Final gathering : Accuracy = 800 / Max rad = 0.05 / Min rad = 0.025
Rendering : size = 1280 x 984 / time = 1 h 51 mins / 4 threads / computer = dual PIII 733 (256 Mb Ram) + dual PII xeon 450 (512 Mb Ram)

 

Despite the fact that the renderings are a bit cleaner than previous tests, here are the kind of artifacts I still can't get rid of :

 

 

Of course, I tried to put much more photons, to use different values for the GI radius, etc... but in general, the only result is much longer rendering time without any significant improvement of quality.
So, Luc-Eric Rousseau, people at Softimage or Mental Images, any new clever tips to handle Global illumination with clean results?

Arnold's "One light only" movie : a simple interior lighting situation.

 

Another simple test was to mimic the little Arnold movie scene : a simple room with light entering through an opened door.
Here again and this is really the simpler situation you can try (almost a "Cornell" box), I didn't manage to get satisfying results (that is, no artifacts, no noise, etc...) even with unaffordable rendering times. Moreover, the lighting behavior in the Arnold rendering seems to be more realistic.

 

 

Arnold movie "One light only" snapshot

 

Does it mean that Mental Ray isn't able to handle correctly a simple "Cornell" Box?
Come on guys, let's look into that deeply! We must find a reliable method to render such simple things...

 

Here are the parameters :
Global illumination : Spotlight intensity = 8000 / Exponent = 2 / 1000 photons / Photon depth = 3, 3, 3 / Accuracy = 200 / Radius = 6
Rendering : size = as it is / 2 threads / computer = dual PIII 733 (256 Mb Ram)