Author Topic: Let's share noise limits!  (Read 34555 times)

2016-05-22, 19:40:36
Reply #30

Dippndots

  • Active Users
  • **
  • Posts: 296
  • Alex Fagan Co-Founder at The Faction
    • View Profile
    • The Faction
Steyin, if you set your render time to 9min and turn on denoising, what does that look like compared to 10min? Surely it looks better.

2016-05-22, 20:51:02
Reply #31

Juraj

  • Active Users
  • **
  • Posts: 4743
    • View Profile
    • studio website
Steyin, if you set your render time to 9min and turn on denoising, what does that look like compared to 10min? Surely it looks better.

Of course it's this.

Your 10 minute render has much higher noise ratio (more than 10percent) than the latter with 30 minutes (but 10perc. noise ratio is still very high, very noisy). But there is no relationship between that time and threshold.
If you had written down the noise ratio that you had after 10 minute first render, than used the exact noise ratio as limit for the next one, it would be done again in 10 minutes. And they would look identical.

Using any kind of limit, has no effect on actual performance.

I'm finding that a smaller time limit with denosing is producing an image that is noisier than just rendering for the 10 minutes without denoising (in the case of the 3000px).

This is almost impossible, because the overhead for additional passes needed for denoising is rather small. It can produce 'ugly' result (because high noise would produce artifacts and loss of details), but not 'noiser' (noiser=higher noise ratio).

Steyin, aren't you completely just confusing the hell out of it ? Denoising, has nothing to do with Noise threshold limit.  And 10 minute render against 10 minute render with 1 minute denoise will look dramatically better in favour of denoised, with exception of quality loss if noise ratio was too high. But if it was too high to begin with, than your non-denoised test would be ugly and noisy as well.

My 7680px (8k) renders take 8 hours on 40 cores. With denoising, I can denoise after 2 hours (where I have reached about 2.5 +/- noise ratio) and get literally almost the same result, minus some bump map loss.
The same goes for my 2560px tests, I just don't have any numbers from head right now.

Can you post screenshots with framebuffer Stats page open ? Maybe you encountered a bug that needs solving but surely looks like something more trivial.

« Last Edit: 2016-05-22, 21:04:51 by Juraj_Talcik »
Please follow my new Instagram for latest projects, tips&tricks, short video tutorials and free models
Behance  Probably best updated portfolio of my work
lysfaere.com Please check the new stuff!

2016-05-23, 00:50:04
Reply #32

steyin

  • Active Users
  • **
  • Posts: 375
  • BALLS
    • View Profile
    • Instagram Page
Steyin, if you set your render time to 9min and turn on denoising, what does that look like compared to 10min? Surely it looks better.

Of course it's this.

Your 10 minute render has much higher noise ratio (more than 10percent) than the latter with 30 minutes (but 10perc. noise ratio is still very high, very noisy). But there is no relationship between that time and threshold.
If you had written down the noise ratio that you had after 10 minute first render, than used the exact noise ratio as limit for the next one, it would be done again in 10 minutes. And they would look identical.

Using any kind of limit, has no effect on actual performance.

I'm finding that a smaller time limit with denosing is producing an image that is noisier than just rendering for the 10 minutes without denoising (in the case of the 3000px).

This is almost impossible, because the overhead for additional passes needed for denoising is rather small. It can produce 'ugly' result (because high noise would produce artifacts and loss of details), but not 'noiser' (noiser=higher noise ratio).

Steyin, aren't you completely just confusing the hell out of it ? Denoising, has nothing to do with Noise threshold limit.  And 10 minute render against 10 minute render with 1 minute denoise will look dramatically better in favour of denoised, with exception of quality loss if noise ratio was too high. But if it was too high to begin with, than your non-denoised test would be ugly and noisy as well.

My 7680px (8k) renders take 8 hours on 40 cores. With denoising, I can denoise after 2 hours (where I have reached about 2.5 +/- noise ratio) and get literally almost the same result, minus some bump map loss.
The same goes for my 2560px tests, I just don't have any numbers from head right now.

Can you post screenshots with framebuffer Stats page open ? Maybe you encountered a bug that needs solving but surely looks like something more trivial.

I'm sure I'm overthinking it, just gotta get used to what is the right way to go. I tried the same settings (10% limit, .6 amount) for a 4000px render on another job (which I would render for 4 hours without denosing) and the image with denoising finished in about 30 minutes, almost looking identical to the non denoised one, so obviously I'm doing something wrong on the other job. Will have to fiddle some more as the one I'm having issues with is at work, so it'll have to wait until tomorrow.

**EDIT**

After some more tinkering, I have got my previews down to 5-7 minutes. I lowered my resolution to 2400px, and realized I had my GI sampling balance still set higher for final renders (d'oh), so I lowered it back to 16. Left noise limit at 10% and amount at .6, ended up with a preview that had 14.1% noise and took 5 mins 24 secs, which looked good enough for test prints. I went back and increased GI to 24 samples and the render took 7 mins 15 secs, so all is good now.
« Last Edit: 2016-05-23, 20:50:35 by steyin »

2019-07-03, 12:43:55
Reply #33

fabio81

  • Active Users
  • **
  • Posts: 444
    • View Profile
My 7680px (8k) renders take 8 hours on 40 cores

Hi Juraj,
I know it's been a long time and that there have been new big improvements with the latest versions of Corona, but can I ask you what noise threshold useful for rendering 8k?
I'm doing rendering from 5k and I'm setting 6% noise limit, does it seem acceptable or could I get even 7% or decrease?
Thank you

2019-07-03, 14:33:22
Reply #34

Juraj

  • Active Users
  • **
  • Posts: 4743
    • View Profile
    • studio website
I find the noise threshold to be very unrealiable as it doesn't account for texture detail well imho.

If I have an office scene that's 40perc. of image covered in that classic noisy office carpet, the image could be absolutely crystal clean and the Noise would still say 4perc.
But if I have clean white space with food flooring, even 3perc. could look still noisy to me, in bad way.

And worst, if I have scene with large white negative space, like white-sweep studio product scene setup in shadowcatcher material, the Noise Threshold would say 0.8 (!!) noise ratio despite the scene only getting 25 passes and AA still looking terrible.
So I don't use noise threshold at all.

Recently, the new noise filter ("HQ Filter") makes noise looks much more pleasing to my eye so higher threshold is now lot more acceptable.

But I don't have any fixed numbers, I go by how it looks to me. I make some crops and see what they look and then set finals for per-passes limit (i.e 200 passes at 16 GI/AA).
Please follow my new Instagram for latest projects, tips&tricks, short video tutorials and free models
Behance  Probably best updated portfolio of my work
lysfaere.com Please check the new stuff!

2019-07-03, 14:41:45
Reply #35

fabio81

  • Active Users
  • **
  • Posts: 444
    • View Profile
thank you for your answer Juraj,

but 200 passes for an 8k are not many?
this is the image where I am working (5000px x 3300px) and I set 6% noise. I'm rendering in a dual 2697V2 but soon I'll do it in a Ryzen 2990wx 32core. I think 6% goes just fine

2019-07-03, 14:48:01
Reply #36

Juraj

  • Active Users
  • **
  • Posts: 4743
    • View Profile
    • studio website
I said that just as random example. I always strive to have perfect AA and for that I need 100 passes...of anything. Whether it is 100 passes as 16 GI/AA or 8,..32.. depends on particular scene. My interiors never have AA issues so I use 16 or 32 GI/AA, for my studio scenes I use 8...or even 4 when I do a lot of Optix denoising so I have the fastest feedback.

But sometimes I need like 500 passes at even 16 GI/AA if the scene is really really problematic with both AA noise (DOF) and and complex GI & Direct light.

Some interiors I do take 1 hour even for 8k rendering, and some 20 hours. The more bounced the light is around complex materials the worse it is...I really go by eye.
Please follow my new Instagram for latest projects, tips&tricks, short video tutorials and free models
Behance  Probably best updated portfolio of my work
lysfaere.com Please check the new stuff!

2019-07-03, 14:49:42
Reply #37

fabio81

  • Active Users
  • **
  • Posts: 444
    • View Profile
thank you very much for the explanation :D

2019-08-28, 20:55:12
Reply #38

cgbeast

  • Active Users
  • **
  • Posts: 58
    • View Profile
Noise level limit set to 5 = 5 percent.

2019-08-30, 03:47:22
Reply #39

cgbeast

  • Active Users
  • **
  • Posts: 58
    • View Profile
Wow...20 hours.  That's a long time these days.  I'm used to rendering 4k images between 30-40 min with denoiser. Max 1.5  hours. So far with 5 percent I've been getting 1:45 on a single machine.

2019-08-30, 18:40:17
Reply #40

serch

  • Active Users
  • **
  • Posts: 39
    • View Profile
Wow...20 hours.  That's a long time these days.  I'm used to rendering 4k images between 30-40 min with denoiser. Max 1.5  hours. So far with 5 percent I've been getting 1:45 on a single machine.

What are your workstation specs?

2019-09-04, 19:08:55
Reply #41

cgbeast

  • Active Users
  • **
  • Posts: 58
    • View Profile
Dual Xeon 3.00ghz
128 gb of ram
GTX 1080 gpu