Open Image Denoise (OIDN) 1.4

[Paul Oblomov]

Here is an Intel and two vray images. How can any of these be confused ?

Vray denoiser always has the same artifacts, which I can't even imagine how you can use in production.

[muoto]

can you specify which RE you need ? i can send you over EXR's (just the ones needed by the denoiser?)

[^Lele^]

Thanks for the info.

Vray denoiser always has the same artifacts, which I can't even imagine how you can use in production.

Via pre-filtering of fireflies.
I can share you a cheap nuke gizmo i did for this, if you want.
ODIN pre-scales HDR images, which we don't.

[^Lele^]

can you specify which RE you need ? i can send you over EXR's (just the ones needed by the denoiser?)

The image in the VFB ought to contain all i need. It doesn't matter if you send me more data, so long as the relevant data is in there.

[Paul Oblomov]

Well, I see hell of a lot artifacts. It would be better for vray to prescale then....

[^Lele^]

Well, I see hell of a lot artifacts. It would be better for vray to prescale then....

You have choices, I'm happy to hear ODIN suits your case best.
That our denoiser isn't fit for your tasks is an undebatable truism (as it's yours to decide.), but that it's generally not usable in production is a comment you're obviously not qualified to make.

[Paul Oblomov]

Heh, denoise it then

[Joelaff]

These samples you guys posted are with the 1.3 Intel engine already included with VRay 5, right?

[Pixelcon]

Here is a quick example from my end. Clearly the intel denoiser is superior when it comes to this particular part of the scene which is glass. Vray denoiser just washes away a lot of the detail for some reason.
IR/LC (however BF/LC produces more or less the same result). Default denoiser settings for both. 200% crop.

[^Lele^]

For our denoiser, you need to set refractive shaders to "affect all channels" or the denoising will not work.

[Pixelcon]

Ahhh that's a problem for me. I need it set to colour only, otherwise some of my photoshop tricks do not work with the reflection channel (which becomes infected with refracted reflections when set to affect all channels).

[^Lele^]

Example denoise (base image rendered to 0.25 noise threshold):
OIDN does a good job at *some* of the detail, when under direct viewing condition (see the burl on the roof).
Notice what happens instead when the seen stuff is in a reflection (left side of the image.), and how the stuff behind glass goes all "painterly", for lack of a better word.
It also removes an octave of detail from the sheet and pillows.

[^Lele^]

For glasses, unless the denoiser is "naive" (i.e. it works with only RGB data, no normals or albedo), there is no way around having channels go through glass.
Taste (i.e. who's doing a better job...) is surely debatable, quality (...at doing the job wrong) is not: OIDN results are just as horrible as ours, if in a different way.
OIDN does its ML-y swirly thing, so it's good so long as the detail can be left chaotic (not a straight line in sight, in other words), while the NVidia one adds some detail it thinks should be there (while it's not in the converged image.), possibly doing the best job of them all under these trying circumstances.
Also, none of this is temporally stable (and how could it, without the benefit of albedo and normals.) for OIDN or the NVidia denoiser, while it *can* be made temporally stable with ours, via multi-frame denoising (i can't quite see why bother, for these kind of results, alas.).
Notice the latest Optix 3.7 adds temporal stability as optional (not sure about the quality), while OIDN has yet to announce a temporally-stable version (while they announced they're working on it).

[^Lele^]

Attached, a Contact Sheet for the three renders above, against a ground truth (rendered at 0.0025 N.T.).
While it's obvious our blurry results are "wrong", numbers tell the story i mentioned above: the other two also make wild guesses as to what should be there, only marginally getting it closer to the ground truth (by Average. V-ray seemingly keeps the sharpest details better, see the Max value.).

The guessing would be even fine if one liked how it looked, and one didn't animate, but if those two conditions weren't met, it'd still be quite unusable.

[Pixelcon]

The update has not yet been implemented in V-Ray, however, it will be. What is the purpose of your question, if I may ask?

Any ETA on this?