Instance test : Memory saving versus Render time.

[ROUBAL]

Hello, I have done new tests about instances used for classical architecture details.

I wanted to see the ratio Memory used/Time to prepare the setup/Render time.

I have tested on one of my detailed buildings.

The first images below shows the Memory use and render speed with a very simple nude structure as initial obj file on which I added all architecture details that can be repeated at least twice, by instances. It required many exports for obj details and csv scatter files, and so much work and time to build the setup in the graph editor.

The second image shows the same building (materials not adjusted) with all details on it, just cut in half and instanced with a symetrical instance.

The render speed is around 7x faster ! The memory use is not much different : 88.0 MB with high use of instances, versus 103.4 MB with just the building cut in two parts. Considering that the central part around the main door has not been actually cut in two and is repeated, there is no much difference.

The third image shows the difference of complexity in the two setups.

If you think a moment that the whole setup may have to be repeated, a high use of instances can lead to a very crowded screen and much work for no significative gain in term of memory and a huge loss of computing power.

So, the test shows that instances are interesting for duplication of complex structures (detailed buildings, cars, trees) as well as particle objects that couldn't be placed by hand, like grass.

The high use of instances for many objects with intermediary level of detail leads to much work and long render time without significative memory gain.

[LudovicRouy]

But there's no materials on windows on the second render.

With textures render still be more faster ?

[ROUBAL]

Materials are there. They are just not adjusted, but you are right, some procedural textures are missing, and have an effect on render time, but globally, there are many less materials, as instancing architecture details requires to duplicate their materials as well. So, the render speed may be not 7x faster, but in my last test at least 4.5x with less work and complexity.

[bepeg4d]

hi,
you can do a clay render for a fast comparison
why the difference in polycount is so narrow?
i would expect a big difference from over 700 meshes versus two
ciao beppe

[ROUBAL]

The polycount is amost the same. I think that this is normal. In the highly instanced setup, the polygons are distributed in many meshes, but they represent the same structure as the building made of only two parts.

I am not yet sure about the method I will use in my final work. I want to keep the maximum flexibility and simplicity without loosing too much memory space or speed (as I would like to keep the possibility of animation later).

I think that the best compromise would be to build modular buildings with not too small parts that can be mixed to build several kind of buildings.

An interesting possibility is offered by the ability of making materials transparent at will. So, it is possible to put several variations of the same detail in given mesh.

For example, on the image below, I have made transparent the rectangular stones on the yellow building. I could have made other shapes of stones in the same mesh.

So, the amount of polygons of one mesh is higher, but less meshes are required. The render time is lower, and it is possible to adjust the final look of the buildings in Octane instead of Blender.