LuxCoreRender Forums

I have a few problems with Tiled Path rendering and need some help or explanations.

Why are the noise levels different when using the Multipass option, which is enabled by default? I used the default settings. Here's an image showing the issue.


What are the best settings to avoid this?

Are you using the adaptive rendering option ? Or are you in the middle of an additional pass ?

I'm not sure what the "adaptive rendering option" is, but this is the rendering saved out after being finished. It's also weird that the more complex area shows less noise than the "easier" areas.

Post a test scene.

I will have to create one - can't post this one.

When I display the tiles stats, the problem becomes clear. Some tiles are stopped after 2 passes (128 samples) while other ones go on (in the image below some are in the 30s already - over 2000 samples). This results in the buckets with more passes to be cleaner and the borders of the tiles become visible. Not sure how to make sense of this system. It may work in scenes that have average material situations. But when there are refractive objects next to simpler ones, you will run into problems - with the default settings at least. I will play with the settings a bit more. Maybe I can find a middle way that works.

Dade wrote: Tue Mar 16, 2021 10:22 am Are you using the adaptive rendering option ?

Yes, you are.

Anyway. At least I know now why Tiled Path is hidden away.

My experience is that the noise-detection algorithm used in tiled path stops tiles at very uneven noise-levels - at least judged by a human eye. The algorithm seems to judge these to be at the same noise level, but they clearly don't look it to a human.

You could try to lower the convergence threshold to mitigate this problem, or disable adaptive rendering altogether.

This system is very chaotic and unless you know what values to use, I would recommend to set "Convergence Threshold" to 0 by default, by which it's basically deactivated. Everything else is pure guesswork. Even using the same Multipass values with different AA Sample values, leads to different results. This is basically a trade of time for acceptable errors. I will do a few more tests to see if there is an average setting that will still be OK, but in the end I will probably leave it at 0.

There also seems to be a bug somewhere. In my big scene I had it running for over an hour and almost all tiles were converged, when suddenly all converged tiles were gone (all green markers disappeared) and they were considered again and rerendered. Might be a bug or some method that makes this even more complicated to understand. From what I have seen in smaller test renderings, once a tile is converged, it stays so til the end.