Denoiser testers wanted

[vlado]

Am I missing something?

A newer build, I think, is what you are missing... some of the recent nightlies did crash because of some stupid bugs.

Best regards,
Vlado

[Rick Eloy]

Really? Hum... I just got that one from the support. Any chance of me getting a more recent build?

[vlado]

Yes, you can email me to vlado@chaosgroup.com and I'll see what can be done about it.

Best regards,
Vlado

[Rick Eloy]

Done. Thanks for the help, Vlado.

[vlado]

Here is a quick test of the standalone denoiser that Lele did for me:
https://ftp.chaosgroup.com/vlado/ima..._test_pngs.rar

The frames were rendered at noise threshold 0.1, with default light cache settings (single-frame mode, not fly-through) so the raw results are crap as expected, but they rendered in about 8 minutes per frame. The frames were then passed through the command-line denoiser, with per-element 3-frames cross-denoising, strength 5 and radius 10.

Best regards,
Vlado

[super gnu]

that is rather impressive. i cant see any artifacts at all in the denoised one. RT gpu or straight vray?

love to see a longer sequence.

[vlado]

RT gpu or straight vray?

It's the regular CPU V-Ray.

love to see a longer sequence.

Yes, will let something run for a bit longer.

Best regards,
Vlado

[super gnu]

id also like to see some fine fur or some dense foliage in there to test it harder! - i guess we can also test ourselves

[stilgarna]

Here is a quick test of the standalone denoiser that Lele did for me:
https://ftp.chaosgroup.com/vlado/ima..._test_pngs.rar

The frames were rendered at noise threshold 0.1, with default light cache settings (single-frame mode, not fly-through) so the raw results are crap as expected, but they rendered in about 8 minutes per frame. The frames were then passed through the command-line denoiser, with per-element 3-frames cross-denoising, strength 5 and radius 10.

Best regards,
Vlado

Very impressive !

[super gnu]

been thinking about the animation denoise.

it seems obvious that the temporal element to the denoising helps a lot, the results seem much better than for a still with the same noise level.. however this leads be to a concern:

given it uses, in this case, 3 frames..

what happens if a noisy element appears suddenly in shot? is it more noisy for the first frame or 2 until the denoiser has 3 frames to compare for the surface in question?

obviously in most cases motion blur on such a fast moving object might help, as would the dynamic nature of the movement.


however i can picture a scene where you pass through a doorway, close to the frame. the room beyond would be revealed quite rapidly, and due to the distance to the room surfaces, they would not be very blurred ( the doorframe would be the fast moving element in shot)

tl-dr: does the denoising vary in quality depending on scene motion?


one other question re interpolation frames: does increasing the number always improve the quality, is it shot dependent, or is it like the interpolation frames in irradiance map where too many frames leads to artifacts in the result?

[vlado]

what happens if a noisy element appears suddenly in shot? is it more noisy for the first frame or 2 until the denoiser has 3 frames to compare for the surface in question?

Well, the first time it appears, the denoiser already has two frames to work with (the current one where the object first appears, and the next one when it is already in the frame). Plus, you can adjust how many frames to consider for denoising (though this increases processing time linearly).

tl-dr: does the denoising vary in quality depending on scene motion?

You'll have to tell me that; we haven't done enough tests yet. I think it will be ok most of the time.

one other question re interpolation frames: does increasing the number always improve the quality, is it shot dependent, or is it like the interpolation frames in irradiance map where too many frames leads to artifacts in the result?

For the first part - yes, more frames improve the quality as the denoiser has more information. For the second question - I don't think there will be artifacts due to the way the denoiser works, which is different from the irradiance map. The denoiser looks for similar blocks of pixels to do the blending, f.e. it will be able to track a slow moving edge across the frames and blend it for the final result. However it will not blend patches that are too different.

Best regards,
Vlado

[super gnu]

ha.. youre right about it having frames to work with.. i had basically forgotten that this is a post-process :P

[^Lele^]

https://www.linkedin.com/pulse/denoi...manuele-lecchi

A little more explanation, and the newer version of the raw render and denoised sequence: with a fixed LC, and the same noise threshold.

[Moriah]

Wow, the denoised version looks amazing, i specially liked the fact that it kept the texture detail on the tree texture and carpet, since it was an issue i had last week with high detail textures. Good job guys, can't wait to try this new version!

[alex.nozadze]

to activate denoising I have to add only render element ?