First couple of tests for review


#21

I temporarily retract that claim, I made a mistake in my measurements. I’ll do them again tomorrow morning.

EDIT: Rendering times with 16 samples/pixel and 1 pass:

  • 8192x4096 HDR map: 94 seconds, out of which almost 6 seconds are spent building the HDR sampling structure, which is a fixed cost and should be excluded, so about 88 seconds
  • 512x256 HDR map: 82 seconds

To conclude, the resolution of the HDR map has a rather small impact on rendering times.


#22

I do like the speed of GPU’s for sure, but I like the idea of a pure CPU renderer. It has flexibility, especially with memory. Having both is 1st prize. :slight_smile:


#23

This is called hybrid rendering and various render support it. Vray has a robust version of it and also Cycles with Blender 2.8 or with the daily builds of 2.79 can use both cpu & gpu together.

To have an idea what you can expect, look at the scene from Marc Gibault posted in the McNeal (Rhino) forum ( https://discourse.mcneel.com/t/vray-3-6-gpu-cpu/54044/19 )

You can see that for this specific scene, using the denoiser brought the largest reduction in render time.


#24

Wow, that is impressive. I have found the denoiser in Cycles has made a massive difference. How developed is the denoiser in Appleseed? I have tried it, but it feels very aggressive with final output having quite a bit of jpeg style artifacts. This is most likely me not knowing how to use the setting properly.


#25

So, I had a play with the settings and the denoiser. I got the rendering time down from over an hour to 12 minutes. It is not perfect, with a few fireflies about and splotchiness but the denoiser has a massive effect.


#26

The denoiser used in appleseed is BCD: https://github.com/superboubek/bcd

BCD is a production denoiser designed to remove the last bits of noise from almost converged images. It will produce artifacts when used on very noisy images.

If I remember correctly (@est please correct me), we still need to tweak internal settings in order to reduce artifacts. We’ve planned to but haven’t got around to do it yet.

BCD is also very problematic when it comes to bright pixels near dark ones, as can be seen on your image around the floor light.


#27

The implemented denoiser works essentially as a smart filter in post-processing by using statistics about the pixels sample distribution. To do this efficiently, the algorithm needs enough data (samples), to get reliably statistical information. As @franz mentioned, it will produce artifacts when used on very noisy images.

Below is a test scene rendered with a very low sample count of 16 samples/pixel. The denoiser removed efficiently the grainy noise (right part of the image). The same scene, even with 10 times the samples (160 spp), still shows some grainy noise. On the down side is that some fine structures in textures can get blurred out (here visible on the fine scratches on the floor)
The settings I used for the denoiser are: Patch Distance = 0.8, Denoise Scales = 4, Prefilter Spikes = off


#28

Cool, thanks for that info guys. Very useful