This scene from La La Land gave us the idea that music can be really good at representing chaos in harmony, so we tried to create a scene that captures that. For the chaos we imagined colored instruments flying around, and for the harmony part, we thought about what probably symbolizes best harmony in music, the conductor.
This was our draft of the idea, but we quickly realized that having a human in the scene was not a good idea so we settled with having a shadow represent the conductor
We chose to implement the clearcoat BSDF to simulate the coating of some instruments, motion blur to add chaos to the scene, normal mapping to make wood textures on the ground and on the instruments look better, denoising allows us to reduce the amount of samples per pixel and rough dielectric allows us to make see-through instruments to increase the amount of different instruments we have.
1920x1080
28,631,325,714
1024
10
5 hours
All 3D models and textures used are taken from Blenderkit and Polyhaven
Violin
Notes
Guitarre
Trombone
Tambourine
Recorder
Piano
Cello
Trumpet
Symbals
Baton
Hands
Bookshelf
Carpet
Floor
but with wood texture of the bookshelf
Pillar
Top
Window
Bottom
Window
Backround Image
Wall
mateial
Painting
1
Painting
2
Painting
3
Coffe
Table
Flower
1
Flower
2
Sofa
Office
Chair
Pillar
with
Texture
Paper
Pile
Flower
Pot Stand
Template of this page by Ahmed Essa and other template used by Tizian Zeltner
When rendering our first instrument, we quickly realized how bad they looked. But because we loved our idea and don't want the Render to look ugly, we needed a solution. This solution was our Clearcoat. For this, we needed to extend our Principled BSDF, as it could only represent diffuse and metallic instruments.
One challenge about implementing the clearcoat in our Principled BSDF was the calculation of the selection probability and in general, how to decide if we sample Diffuse, Metallic or Clearcoat. But we figured it out and now our instruments don't look bad anymore; they actually look very nice.
Unfortunately, due to Monte Carlo integration, our images contain some noise when we render them, one way to reduce noise could be to drastically increase the amount of samples per pixel but that would come with an increase in computations and would not be scalable, and our render time is definetly long enough.
So instead we use the OpenImageDenoise library from Intel that essentially smooths out the noise. We give 3 images as input to that library, the noisy image that we render normally, the normals image, that contains the information about which direction the normal vector has for each pixel and the albedo image, which shows the colors of the elements without lighting having any effects.
Noisy image
Albedo image
In this image the sphere does not show up because the background has the same color as the sphere and there is no shadow or light to make the sphere stand out
Normals image
Here all surfaces that point in the same direction have the same color, eg. the ground and the top of the sphere both point up and are both purple
Result image
Normal Mapping on Rectangles
Normal Mapping on Spheres
Because just flying instruments are boring, we thought we can have colorful instruments. But of course colorful instruments are also not enough. So we thought instruments made out of glass would be a great addition to our colorful chaos. Maybe we could also play a bit with colored light, which then will be reflected in our glass instruments. For realizing good-looking glass, which also can be a bit rough, we needed to implement a Rough Dielectric. To test this we create three test scenes. One with glass panes, one with glass spheres, and one with glass apes, and all with three values of roughness.
Rough Dielectric on Rectangles
Rough Dielectric on Spheres
Rough Dielectric on Meshes
Motion Blur allows us to add movement to certain objects of the scene, for this, every
ray
gets a random time value between 0 and 1, and out shapes get a motion vector. At time 0
every shape is in its default spot, at time x the shape is located at position: default
+
motion * x. Having a random time for each pixel allows us to integrate over the time,
which
gives us the motion blur effect.
A challenge we faced during the implementation was that the bounding boxes for object also needs to be adapted to include the object at all points in time which worsens performance.
One downside with this implementation is that for large motions or small sample sizes
the
odds of a shape being at a specific place are low, so the motion blur can be hidden,
there
are several ways to fix this issue, either by increasing the amount of samples per
pixel, or
by adding an acceleration value to the motion, which would increase the amount of time
an
element stays at a specific spot (at the beginning or the end) and decreases the motion
blur
on the way, but our scene only required small movements so we did not implement it.
Now if we integrate over the time
