**Improving material system**
In the [previous post](../part-3-2/index.html) we added a multiple bounces to our ray-tracer. We introduced
importance sampling, which helps us to estimate a rendering equation more precisely and implemented global
illumination in the scene with diffuse materials. Now lets make our material system more explicit to be able to
add new materials to the scenes.
As always [source code](http://github.com/sergeyreznik/metal-ray-tracer/tree/part-3)
for this post is available on my GitHub page.
Explicit materials
======================================================================
In the previous post we've added multiple bounces and implemented global illumination. On each bounce we scaled ray's
throughput value by material's color, thanks to the properties of the diffuse material. But in order to have various
materials we want code to be more generic and not dependent of the specific material properties. Generic code would look like:
But since BSDF and PDF depends only on material, and \(cos(\theta)\) could be cancelled within BSDF or PDF, idealy
we would write something like that:
So, in general case we need to have a function, which will sample material, knowing the input direction,
the output direction and the normal in a given point. Such function should return BSDF value, PDF value and combined
value (bsdf_weight) which we will plug directly into the ray's throughput computation. Let's make a structure which
will hold output of such function and declare function itself:
We will use \(\omega_i\) for the input direction and \(\omega_o\) for the output direction. Also we need to store
material type in the `Material` structure. And looks like we are ready to implement this function for diffuse BSDF and
plug it into our ray-tracer. Let's assume that diffuse BSDF will be used by default. We will also compute
\(\cos(\theta)\) inside this function - this will give us an ability to output weight, which will not
contain all cancelled terms:
Generating next bounces
======================================================================
In the main ray-tracing shader we will need to do two things:
1. generate next rays depending on the current material;
2. update ray's throughput according to the current material.
But our function can only sample BSDF and PDF knowing input and output directions. We need to add another function,
which will actually generate an output direction from two random variables (as discussed in the previous post)
and then sample the material:
Having such function we are ready to plug it into the our main ray-tracing shader:
One more application of the `sampleMaterial` function - would be use it in the explicit light sampling. As you
should remember from the second post of this series, we explicitly sampling light sources. To do it we generating rays
towards randomly chosen points on the light sources and if we actually hit emitters - we are adding radiance to our
primary rays. And now we can sample material in this direction in order to determing how much incoming light
contributes to the given point:
Adding mirror (conductor) material
======================================================================
Now, everything seems to be ready to add another material. Let's start from adding a mirrors to our scene.
This is perfectly smooth surfaces that reflects each incoming ray into the one specific direction. Therefore,
BSDF of such material will equal to `1.0` in the direction of reflected ray and `0.0` elsewhere. So will be a PDF and
their ratio - BSDF / PDF will only be valid and equal to one in the mirrored direction and zero elsewhere.
Generating a new direction for such BRDF (this is actually BRDF, not BSDF, because it is not scattering light) is quite
easy, let's add add another case to out function which generates rays:
Sampling mirror material is also straightforward - we should check if given output direction is close to mirror, and
return 1.0 everywhere if it is:
The most interesting begins when it comes to sampling lights. Let's add another scene, a box with spheres and water
surface inside. If all materials will be diffuse, scene will look like:
![](images/pic-1.png)
But if we will change spheres and water to be mirrors, we will see this one:
![](images/pic-2.png)
Light sources are black. The trouble in the explicit light sampling. Mirror BRDF and PDF are only valid in one
direction, and probability of generating this direction when we sampling light source is zero, therefore we never
actually makes any rays toward light sources that will contribute to the mirror BRDF.
Solution for this would be to not sample light sources explicitly, but add radiance to the rays when we hit
emitters during the bounces. This is the perfect case for the mirror BRDF - we are generating reflected ray and if it
hits the emitter, it contributes to the total radiance along this ray. We can easily extend our ray-tracing shader to
account emitters:
Let's also disable explicit light sampling and see what will happen:
![](images/pic-3.png)
Light sources are now visible, but look at diffuse walls - they are very noisy even after 700+ samples.
Let's limit ray-tracing to 16 samples and compare diffuse surfaces with bsdf sampling and explicit light sampling:
![](images/pic-4.png)![](images/pic-5.png)
Difference is amazing. This happens because we are not generating enough samples to integrate over whole hemisphere,
and therefore most information about lighting is missing on the diffuse surfaces. How can we deal with this?
Adding rough conductor material
======================================================================
Before we deal with the problem described above, let's add another material - rough conductor. This is commonly used
material, which reflects light in a cone, specified by it's parameter - roughness. The less roughness is - the more
mirror like material becomes, and when roughness value is large - material reflects light in all directions.
We going to use widely used microfacet material with GGX distribution. Since this is really one of the most used materials,
I will just provide a pieces of code here:
Sampling material's properties (inside `sampleMaterial` function):
`ggxNormalDistribution` and `ggxVisibilityTerm` are also well known, so I don't see any sense to provide them here.
`fresnelConductor` could be hardcoded to return `1.0` now, and then updated with something very physically-based :)
To generate a new direction according to this material we will use:
Let's plug this material into the scene and see how it looks like with explicit light sampling and bsdf sampling:
![](images/pic-6.png)![](images/pic-7.png)
Difference is not that obvious on this scene. But Eric Veach [in 1995](https://sites.fas.harvard.edu/~cs278/papers/veach.pdf)
created a scene, that represents mentioned problem quite well. There are four light sources, each other in larger but
has less intensity, also there are four bars and each other is more rough.
Let's recreate this scene and render it with light sampling and bsdf sampling:
![](images/pic-8.png)![](images/pic-9.png)
On the left image (with explicit light sampling) more smooth bars lacks proper reflections of the largest emitter. This
happens because when we are generating rays toward light sources - we rarely generating them in the mirror direction,
and such rays does not contibute to the final image. On the right image (with bsdf sampling) we generating rays according
to the material's distribution, and therefore almost always hits light sources. This works for almost mirror bars,
but fails for bars with large roughness and diffuse wall behind, because randomly reflected rays does not always hits light sources.
Multiple importance sampling
======================================================================
So how we deal with the problem described above, where not BSDF sampling, nor light sampling converges well? Solution
would be to sample both of them - BSDF and the light sources. But imagine a situation where you sampling light source,
adding some radiance to the ray and then sampling BSDF in the same direction. And again ray will hit same light source
along the same direction, and this light source will contribute to the ray's radiance twice. Seems like we can average
samples in this case (i.e take light sampling result with weight 0.5, and BSDF sampling result with the same weight).
But this will work only when we are sampling BSDF and the ligh sources in exactly one direction. But since there are a lot
of random values around - this rarely happens. The real solution for this problem will be to use **balance heuristic** for weights.
Let's say \(f_p\) is the probability of sampling light source, and \(g_p\) is the probability of sampling BSDF (generating
a new direction for the ray, exactly what we are doing in `sampleMaterial` function). Then weight of the light sampling
would be:
\(weight_l = f_p^2 / (f_p^2 + g_p^2)\)
And the weight of the BSDF sampling contribution would be:
\(weight_b = g_p^2 / (g_p^2 + f_p^2)\)
(read more about Multiple Importance Sampling in the
[Chapter 13.10.1](http://www.pbr-book.org/3ed-2018/Monte_Carlo_Integration/Importance_Sampling.html#MultipleImportanceSampling)
of the **PBRT book**)
Let's plug this to the shaders, and check what will happen:
Also we need to modify our BSDF sampling code in the shader. We have to determine a probability of sampling light
in the direction, generated by BSDF samling (our `sampleMaterial` function), it would be pretty much the same,
as in the light sampling code:
As you may notice - we've added another value to our `Ray` structure - `materialPdf`. This is the value, retured by
`sampleMaterial` function, and it is used only for the multiple importance sampling in the next intersection handling shader.
The result of such modification would be visible immediately. The same scene after 64 samples:
![](images/pic-11.png)
And after 700+ samples:
![](images/pic-12.png)
Diffuse surfaces around small light are being sampled with explicit light sampling, and large emitters are visible in
the almost mirror bar because of BSDF sampling. Everything looks fine, and actually it is, because the same image is
produced by Mitsuba renderer (I've included sample scene):
![](images/pic-13.png)
**Thats it!**
[Next post](../part-4/index.html)
[Return to the index](../index.html)