LuxCoreRender Forums

kintuX wrote: ↑Wed Jun 23, 2021 1:23 pm

B.Y.O.B. wrote: ↑Wed Jun 23, 2021 12:50 pmSwitch what?

For OIDN to use "Shading normal" instead of "Avg." pass.

No, it's not possible to change that. OIDN can only work with the averaged normals.

B.Y.O.B. wrote: ↑Wed Jun 23, 2021 2:35 pm

kintuX wrote: ↑Wed Jun 23, 2021 1:23 pm

B.Y.O.B. wrote: ↑Wed Jun 23, 2021 12:50 pmSwitch what?

For OIDN to use "Shading normal" instead of "Avg." pass.

No, it's not possible to change that. OIDN can only work with the averaged normals.

Oh, so... baad ...no fine details preserved either way, maybe sooner or later but in meantime, it would really be nice to have choices/options juangea mentioned available

It's not bad. I have a feeling you don't understand what the difference between the two pass types is?

Ok, this is what I found.

This is a rendering with 100 spp:


There is clearly a filtering problem (i.e. the moire effect).

Even more samples per pixel will not help, this is a 1000 spp rendering:


However, if I re-enable pixel filtering (note: disabled because of OIDN requirement), even a 100 spp rendering is a lot better:


So I tried to just use ODIN with pixel filter enabled and the result looks a lot better. This is a 100spp rendering (first is raw, second is denoised without pixel filtering, the latest is denoised with pixel filtering):


I would say that latest is a far better denoised result. We clearly need to have pixel filtering enabled, at least on Normal pass.

Unfortunately, because the why pixel filtering works, I can not conditional filtering on some AOV and not on others (I would have to literally render the image multiple times).

At the moment, my conclusion is that we should re-enable pixel filtering for path tracing + ODIN. BiDir should probably stay like it is because, I'm pretty sure, the way pixel filtering work in BiDir will drive OIDN crazy.

Without any doubt the filtered image is way better.

A good test could be to have lit and shadow zones, with a sun, also inside an interior.

Have you tested with the albedo being considered like perfectly clean?

I think that should improve the result too, specially in complex scenes where the cloth patter is lost, like the pictures I showed before.

juangea wrote: ↑Thu Jun 24, 2021 2:19 pm Have you tested with the albedo being considered like perfectly clean?

Albedo is pre-filtered in v2.6 so yes. I have also tried to pre-filter the normal pass to see if it helps but it only increases the "blurriness" of the end result.

No, I mean the mode where it considers the Albedo as noise clean but without prefiltering.

AFAIK 1.4 has three nodes:

- As before
- Prefiltering Albedo
- Consider Albedo as prefiltered (noise free) but without prefiltering it.

In situations with fine detail like this one, but where you don’t have normals, it may be possible that you want Albedo but no bump, you may loose a lot of detail with the prefiltering, because it considers the texture detail as noise.

So what I mean is the third mode, where the Albedo is not prefiltered but is considered as noise free.

I’ve been looking at the API here:

https://github.com/OpenImageDenoise/oid ... readme.pdf

I mean the CleanAUX filter parameter.

B.Y.O.B. wrote: ↑Wed Jun 23, 2021 11:47 pm It's not bad. I have a feeling you don't understand what the difference between the two pass types is?

Yeah, it's good. I was aiming for too fine detail (almost subpixel size).

juangea wrote: ↑Wed Jun 23, 2021 2:04 pm One important thing regarding OIDN 1.4.1:

chrome_FNaHVXsBJW.png

I think we are still on OIDN 1.4.0. Should we do this upgrade