[Mihai]

Right. If you haven't already, use the tech support form to notify them of this bug, it's pretty important...

[seghier]

I will report this bug

[tom]

We have it internally reported.

[seghier]

We have it internally reported.

Good

[Fox2]

Hi all!
I did some tests of 4.1 and relative denoiser, I can definitely see a great potential in it.

Surely I need to go deeper in testing, for the moment I can tell - as downsides:

- It looks like the denoiser is eating away some details (eg. orange slice), even if its complex architecture is supposed to preserve most of them (and it actually does, in most of the cases, as far as I can tell). For this reason, I think it would be great to have some kind of control over the amount of denoising (like denoise radious/strength or similar) prior rendering.

- Some weird artifacts happening on sss surfaces (eg. beginning of monster's tail).

- From the orange juice test (I could not upload the original uncompressed tiffs, too heavy), it can be seen colour banding artifact happening on the limbo/background when the denoise is applied.

- At the moment, the denoise setup is very simple and ergonomic - which is a great thing. To make it even more clear we could avoid the "denoised image path" (does make sense just to save it where all of the other rendered passes are), and maybe put as default the option "Apply denoiser At the End", which seems to give the best results in terms of quality.

- The second pass (when denoise happens at the end of calculations) in some conditions is kinda slow (can take hours actually).

- in 3DS Max plugin, the denoiser tab has to be checked "enable" every time the scene is re-opened, which is kinda annoying.

- EDIT: I forgot one thing. To me, the re-denoise button should be Also put below the "refresh/preview" one, in Maxwell Render interface. Would be more immediate and intuitive.

Gonna keep on testing 4.1 on new personal projects. Great job done so far NL team!

ps. Is there any news about GPU engine development status? When will it reach the same capabilities/features of the CPU one? (Sorry for the ot)

Best,
Riccardo / ramonesFAN



First image is SL 13
Second one is SL 14

[seghier]

I think you need use extra-sampling and give some parts a high sl and use denoiser
I didn't tried that and i hope it work for you or maybe one from maxwell team confirm if it work or not

[luis.hijarrubia]

- The second pass (when denoise happens at the end of calculations) in some conditions is kinda slow (can take hours actually).

Can you explain this a little bit more? I mean, second pass is just a second render with different seed, it should last the same as the first one. Do you find it slower than the first?

[Fox2]

@seghier:

Done in the following test, attached here.

@luis:

I considered the second pass with the denoiser process too (while they are distinct, from what I can see in the documentation). However, I think in the previous tests I misread the log datas, my bad.


Gonna report the log's essentials of this test (Extra SL + denoised SL:10):

First pass (to SL 10, real SL 13.52): 12:03-14:13 / 2hours 10mins
Second pass (to SL2 10, real SL2 11.82): 14:13 -16:19 / 2hours 6mins
Denoise pass (SL 10 - Merged display with SL 13.5227): 16:19:15-16:19:33 / 18secs

The second pass was actually quicker than the first.


What I can show you here, is a strange thing happening at the end of the calculations, when the denoised image is generated.
The den. image looks very "drowsy" and the details blurred...until you press on re-denoise, then the same denoised image is re-calculated and everything looks way better. I thought it would have been a good idea to underline this behaviour, which could led to misleading judgments of the denoiser efficacy. Dunno if this has anything to do with multilight set-up.

ps. luis, could you explain to me why do we need a second render / second pass in the denoise process? So to already obtain a cleaner result before the denoiser happens, with the mxi merging process?

Thanks guys, hope these quick tests will be useful somehow. Looks great tho!

R

[Mihai]

The den. image looks very "drowsy" and the details blurred...until you press on re-denoise, then the same denoised image is re-calculated and everything looks way better. I thought it would have been a good idea to underline this behaviour, which could led to misleading judgments of the denoiser efficacy.

Wow Nice find....that's not exactly a subtle difference. Nasty evil bug there.

About the two render passes, the denoiser looks at differences in noise patterns to better determine what is actually noise and what details are in both passes. Also the reflectance pass helps with this, but there is currently a bug with this and instances which seghier pointed to earlier. The textures don't show up in the reflectance pass for instances.

Great tests btw! Really shows cases where the denoiser can be of tremendous help.

[seghier]

Gonna report the log's essentials of this test (Extra SL + denoised SL:10):

First pass (to SL 10, real SL 13.52): 12:03-14:13 / 2hours 10mins
Second pass (to SL2 10, real SL2 11.82): 14:13 -16:19 / 2hours 6mins
Denoise pass (SL 10 - Merged display with SL 13.5227): 16:19:15-16:19:33 / 18secs

beautiful and clean result ; especially the sss material

[Fox2]

@Mihai

Thanks for the explanation, makes total sense.
Glad if you guys will find these tests useful in any ways!

@seghier

Thanks mate, I'm mainly testing scenes with SSS 'cause - in my experience - it's always been the noisest and hardest thing to render properly.
I remember when back in the days I rendered the same scene (last one I posted) until SL 15 to have a "sort of" clean result, in 10hours. Insane. Was no way near to the denoised version you can see here, obtained in 4hours. Still long calculation time for such a simple scene, but surely it's a big step forward!!

Then, I am still waiting for the GPU engine to make the Real debut in production. When GPU will support all the features we have in CPU, with the denoiser I can see a very bright future for V4!

Great work done so far by the team.
All the best!

R

[T0M0]

Some news about Maxwell betas ?

[Tea_Bag]

Just wanted to bump this - any news on Maxwell Betas?

[Mihai]

The newest update fixes a few issues with the denoiser, but it doesn't mention the very nasty denoising bug reported above - that the result is much better when you press denoise a second time (I also tested this and got a similar result - just pressing denoise again in Maxwell, makes previous very blurry areas much better). I hope this is on the list in the near future?

[Fox2]

Quoting Mihai here,
any news about resolving then denoiser bug mentioned above?

Best!
R