[Bf-committers] Realistic materials question (Farsthary)

[Yves Poissant]

From: "javier loureiro" <derethor at derethor.net>
Sent: Wednesday, February 25, 2009 7:34 AM


> I believe that the main problem with complex BxDF's is how to sample it
> in an efficient way. There is no a way to sample according to a complex
> BxDF (a tabulated one, for example) , so, you sample cosine weighted
> around the normal.

There already exist different techniques for importance sampling BxDFs. 
AFAIK, though, importance sampling tabulated BxDF is not efficiently solved 
yet but given that tabulated BxDF are not easily and/or cheaply accessible 
right now, I would not consider this an important issue for Blender.

However, good importance sampling strategy do have a very important impact 
on performance for non-tabulated BxDF as well. And there are many importance 
sampling strategies and new ones are designed regularly and different 
strategies are better suited to different situations. An implementation in 
Blender should, thus, attempt to include that concept in some more or less 
abstract way. This is a thought I have since I would not know how to design 
and implement an abstract importance concept right now but this looks to me 
like a necessity and have been simmering on my backburner for a while.

But for truely efficient importance sampling, the samples should not only be 
distributed according the the BxDF PDF but also according to the environment 
lighting PDF. Imagine the BxDF PDF is very strong in a direction where the 
environment PDF contributes no light. So the optimal PDF from which samples 
should be drawn is the product of the BxDF PDF and the environment PDF. That 
complexifies the matter a bit more.

> Also, you must define multiple leyers of BxDF for a
> material, with different weights (like 20% diffuse, 60% phong, 20%
> emissive), and "weight maps". The random number generation, and previus
> stratification of samples is fundamental to sample correctly the whole
> material.

>From my experience, good results can be obtained by using a 
multi-dimentional low-discrepancy vector sample generators. Generate one 
sample per path and then use additional dimension for sampling the light, 
the material, the different layers, the bounces, etc. Those additional 
dimensions can be supplied on demand and used for bounces too. There are 
more complex schemes but this one is cheap and works very well.

> Another question: what happens when a material is emissive (like lava)..

An emissive material does emit light. It is a light source.

> how to deal with non realistic light sources, like point lights,
> directional, etc?

There are easy solutions for those. The actual issue is not so much the 
concept of point light or directional light themselves as how their 
properties are defined by the user. For instance, how do you translate a 
light that have been specified with an energy of 1, a half distance of 30 
and a linear attenuation? What would that mean in term of physical 
quantities?

> pd. first time i write here :)

Cool! Welcome to the discussions.

Yves