Open image denoiser - strange effects after the first denoising pass

[Micha]

... the first run looks promising, but the next denoising passes looks strange.

[Micha]

... I found that the Nvidia denoiser shows an equal progress and it looks like the good looking first result based on the LC pass and the later passes are GPU based. Could it be that the denoiser are not optimized for GPU?

[Micha]

... on the long run a progressive interactive CPU run shows strange structures too.

[vlado]

There's quite a bit of noise in those passes so I suppose the denoiser makes up some details. This will get better when the light cache is enabled in interactive mode, I hope.

Best regards,
Vlado

[Micha]

I checked the progress of the Intel open image denoiser during a progressive RTX rendering and the result was like expected, an increasing quality. Very nice. Looks like only the real time calculation and the denoisers doesn't work well together. I hope an improvement is possible.

[Micha]

Here a result at 2560x1440 after ~40s of progressive rendering per LC+RTX. Bigger size, shorter time and a much better result. I'm impressed. The real time progress of my first tests was so much less quality ... .

Since I don't have success to the parameters of the real time rendering I can't test other parameters, but could it be that this parameters needs to be optimized only? Why are ~3min RT looking so much more worse than 30s production render?

[vlado]

As I mentioned above, the reason is the light cache, which is currently disabled for interactive rendering. We are working on this however and in one of the next updates the interactive rendering should match the production one.

Best regards,
Vlado

[Micha]

Great, I'm very curious.

[Micha]

Sorry, I'm back again. I got an idea for a quick test and disabled the LC in production mode - a very good result, also without LC. So, could the results of the RT mode not much better also without LC?

[Micha]

...and one more test. I set a longer render time in production mode and got artefacts also per LC.

The goal of my tests was to find a way to quick render fly throughs. It looks like there is a very limited range of parameters where very nice results can be reached (see post above). It would be great if this result could be used for further optimizations of the code/parameters. For me it looks like in the first seconds of the final passes anything works quite well. Than the noise from the lights increase and starts refining. In this moment the denoiser creates artefacts. Maybe after some minutes this noise of the lights is fine enough for the denoiser. So, I have the choice to render a few seconds or a few minutes, but nothing between.

[Micha]

Finally I found there is a bug with the ceiling lights and the LC (bright artefacts at my last posted image). It's reported as 330-710-7313.

Also the reported simplified scene shows how clean the progressive rendering works (BF) from the start, but the interactive rendering fails more and more over the time (dark denoiser artfacts).

I spend some hours for this issues now and I'm curious what can be found.

(There is also a scene packing crash bug.)

[konstantin_chaos]

Hi Micha,

Finally I found there is a bug with the ceiling lights and the LC (bright artefacts at my last posted image). It's reported as 330-710-7313.

I was suspecting something like that.
Sometimes fireflies and such appear during the later sampling passes and might confuse the denoising algorithms.

And as Vlado suggested - the better the lighting in the initial passes, the better the denoised result would be.
The Light Cache will help a great deal when available during interactive rendering.

Konstantin

[Micha]

konstantin_chaos I'm curious for LC for RT. Some questions stays open:

(1) Why is the fire fly noise in a rectangular area only?



(2) Why is the interactive rendering looking so bad in comparison the the non interactive rendering? See my attachments, both are not using LC. Could it be possible to use the same settings from production for RT? The production setups clear up much faster.

(3) I sent two LC cache files. Can you find the the reason why the final calculation based on the one LC cache looks clean and the other LC cache shows a fireflies noise? Something is different between both, but I don't know what? I would like to get the clean result always, but I don't know what caused the clean look.

[konstantin_chaos]

Hi,

(1) Why is the fire fly noise in a rectangular area only?

I believe this was discussed in another thread and you shared the scene.
I'll test it myself and get back to you.

(2) Why is the interactive rendering looking so bad in comparison the the non interactive rendering? See my attachments, both are not using LC. Could it be possible to use the same settings from production for RT? The production setups clear up much faster.

You are correct - we do use different settings during interactive rendering (otherwise the engine is exactly the same).
You can't set and use the same exact settings that the production engine uses but you can bring them closer by setting the Interactivity to Low.

(3) I sent two LC cache files. Can you find the the reason why the final calculation based on the one LC cache looks clean and the other LC cache shows a fireflies noise? Something is different between both, but I don't know what? I would like to get the clean result always, but I don't know what caused the clean look.

Again, I'll have to go back to the other thread and read you full description of this issue.

Konstantin