Alain Martel <kua### [at] videotronca> wrote:

> > Hi everyone,
> >
> > On the left is a noisy image produced with POV-Ray, and on the right is the same
> > image, but denoised in Blender. It was super easy to set up.
> > (...)
>
> Look like a simple blur operation using a 3x3 block size.

Hi Alain,

I don't think it's that simple... Here's an article about it:
https://www.intel.com/content/www/us/en/developer/articles/technical/intel-employs-ml-to-create-graphics.html#gs.mshzmp


I haven't read through the whole thing, but they used machine learning somehow.
My guess is that they fed a neural network image pairs: one completely converged
(noise-free) image, and the same scene, but really noisy. The system eventually
learned that when the noise looks a certain way, it's meant to turn out a
certain way.

Sam

Post a reply to this message

Here's a striking example of how well Intel's denoiser works. The plastic
material uses backside illumination, reflection, refraction, and a tiny granite
normal for roughness. This kind of material needs at least two radiosity bounces
to look decent, so render times can really go through the roof under normal
circumstances. Observe the reflection highlights. The denoiser was able to make
sense of the clusters of white pixels.

Post a reply to this message

Attachments:
Download 'blender-uberpov-ducting-13m_51s.jpg' (332 KB)

Preview of image 'blender-uberpov-ducting-13m_51s.jpg'

Op 16-1-2023 om 16:22 schreef Samuel B.:
> Here's a striking example of how well Intel's denoiser works. The plastic
> material uses backside illumination, reflection, refraction, and a tiny granite
> normal for roughness. This kind of material needs at least two radiosity bounces
> to look decent, so render times can really go through the roof under normal
> circumstances. Observe the reflection highlights. The denoiser was able to make
> sense of the clusters of white pixels.

This one is particularly impressive, Sam.
-- 
Thomas

Post a reply to this message

A proximity pattern test. This type of proximity method can be quite noisy at
times, depending on the number of samples used. This object got 12 samples and
the scene took 35 seconds to render using low radiosity and focal blur settings.

Just in case any wants it, here's the proximity macro:

/* Proximity pattern macro
 ProxObject = input object
 Radius   = proximity radius
 Samples   = number of priximy samples
 TurbScale  = scale of turbulence
 TurbAmt   = amount of turbulence
 Example: Object_Prox(MyObject, .3, 20, 1000, 100)
*/
#macro Object_Prox(ProxObject, Radius, Samples, TurbScale, TurbAmt)
 pigment_pattern{
  average
  pigment_map{
   // fermat spiral-sphere distribution
   #local Inc = pi * (3 - sqrt(5));
   #local Off = 2 / Samples;
   #for(K, 0, Samples-1)
    #local Y = K * Off - 1 + (Off / 2);
    #local R = sqrt(1 - Y*Y);
    #local Phi = K * Inc;
    [1
     // object pattern with small-scale turbulence
     pigment_pattern{
      object{ProxObject}

      scale TurbScale
      warp{turbulence TurbAmt lambda 1}
      scale 1/TurbScale

      translate <cos(Phi)*R, Y, sin(Phi)*R>*Radius
     }
    ]
   #end
  }
 }
#end

Post a reply to this message

Attachments:
Download 'denoise-prox-test3.jpg' (338 KB)

Preview of image 'denoise-prox-test3.jpg'

Thomas de Groot <tho### [at] degrootorg> wrote:
> Op 16-1-2023 om 16:22 schreef Samuel B.:
> > Here's a striking example of how well Intel's denoiser works. The plastic
> > material uses backside illumination, reflection, refraction, and a tiny granite
> > normal for roughness. (...)

> This one is particularly impressive, Sam.
> --
> Thomas

Hi Thomas,

Yeah, it's amazing how effective it can be. But writing POV script to output
three different images is a bit of a chore (for a few reasons). It would be cool
a future version of POV-Ray supported denoising natively, as a plugin or
something, so that all the difficult stuff was done behind the scenes.

Sam

Post a reply to this message

"Samuel B." <stb### [at] hotmailcom> wrote:
> A proximity pattern test. This type of proximity method can be quite noisy at
> times, depending on the number of samples used. This object got 12 samples and
> the scene took 35 seconds to render using low radiosity and focal blur settings.
>

This is the best example of the denoiser technique that you've posted so far,
IMO. I am amazed that it can discern true noise from your applied ground pattern
that looks *almost like* noise. The foreground detail looks nice and sharp.

It would be interesting to see this same scene denoised but without the original
focal blur-- to see at what point (if any!) in the receding distance the
denoiser might mistake the smaller and smaller actual ground pattern detail for
what it 'perceives' as noise. To see if the denoiser tries to blur it there when
it should not(?)

----
BTW: A few days ago, I downloaded your 2013 file "ToVolume: Object-To-Volume
Conversion and Rendering Process". It also has an interesting and much more
complex 'proximity' file included. I'm sorry to admit that I haven't yet played
around with your amazing proximity pattern :-( I'm still going through that file
to try and understand its workings; much of it is beyond my brain-power, ha.

Post a reply to this message

"Kenneth" <kdw### [at] gmailcom> wrote:

>
> It would be interesting to see this same scene denoised but without the original
> focal blur--

Oops, what I meant to say was, *with* focal blur (and the inherent noise of low
blur samples) but so that the depth-of-field is much wider and not soft-focus
(if that makes sense.) In other words, noisy but not actually blurry-looking.
I'm wondering if the denoiser can distinguish the ground pattern in the
foreground from the same un-blurred smaller pattern in the receding distance.

Post a reply to this message

"Kenneth" <kdw### [at] gmailcom> wrote:
> "Samuel B." <stb### [at] hotmailcom> wrote:
> > A proximity pattern test. (...)
>
> This is the best example of the denoiser technique that you've posted so far,
> IMO. I am amazed that it can discern true noise from your applied ground pattern
> that looks *almost like* noise. The foreground detail looks nice and sharp.
>
> It would be interesting to see this same scene denoised but without the original
> focal blur-- to see at what point (if any!) in the receding distance the
> denoiser might mistake the smaller and smaller actual ground pattern detail for
> what it 'perceives' as noise. To see if the denoiser tries to blur it there when
> it should not(?)
>
> ----
> BTW: A few days ago, I downloaded your 2013 file "ToVolume: Object-To-Volume
> Conversion and Rendering Process". It also has an interesting and much more
> complex 'proximity' file included. I'm sorry to admit that I haven't yet played
> around with your amazing proximity pattern :-( I'm still going through that file
> to try and understand its workings; much of it is beyond my brain-power, ha.

Hi Kenneth,

The denoiser was able to preserve the ground bump largely thanks to the normal
pass I included. It also helps that the normal and albedo passes can take more
camera blur samples than the final pass, since they are faster to render.

Regarding what it would look like with no obvious focal blur, I'm /guessing/
that the denoiser would smooth out everything under a certain size and color
threshold.

Re: ToVolume. I can't remember which type of proximity technique I used for
that. And I would probably be a bit lost myself, opening up that file after all
these years... that tends to happen with old projects ;)

Sam

Post a reply to this message

Op 20/01/2023 om 23:56 schreef Samuel B.:
> "Kenneth" <kdw### [at] gmailcom> wrote:
>> "Samuel B." <stb### [at] hotmailcom> wrote:
>>> A proximity pattern test. (...)
>>
>> This is the best example of the denoiser technique that you've posted so far,
>> IMO. I am amazed that it can discern true noise from your applied ground pattern
>> that looks *almost like* noise. The foreground detail looks nice and sharp.
>>
>> It would be interesting to see this same scene denoised but without the original
>> focal blur-- to see at what point (if any!) in the receding distance the
>> denoiser might mistake the smaller and smaller actual ground pattern detail for
>> what it 'perceives' as noise. To see if the denoiser tries to blur it there when
>> it should not(?)
>>
>> ----
>> BTW: A few days ago, I downloaded your 2013 file "ToVolume: Object-To-Volume
>> Conversion and Rendering Process". It also has an interesting and much more
>> complex 'proximity' file included. I'm sorry to admit that I haven't yet played
>> around with your amazing proximity pattern :-( I'm still going through that file
>> to try and understand its workings; much of it is beyond my brain-power, ha.
> 
> Hi Kenneth,
> 
> The denoiser was able to preserve the ground bump largely thanks to the normal
> pass I included. It also helps that the normal and albedo passes can take more
> camera blur samples than the final pass, since they are faster to render.
> 
> Regarding what it would look like with no obvious focal blur, I'm /guessing/
> that the denoiser would smooth out everything under a certain size and color
> threshold.
> 
> Re: ToVolume. I can't remember which type of proximity technique I used for
> that. And I would probably be a bit lost myself, opening up that file after all
> these years... that tends to happen with old projects ;)
> 
> Sam
> 

There are/were also your "fastProx" and "nestProx" includes for doing 
proximity patterns. It has been a while since I last used them. They 
tended to be /superseded/ by Edouad Poor's "df3prox-0.95" utility in my 
(slight) personal choice/preference ;-)

However, they are a notable part of my large collection of POV-Ray 
utilities created by the users community. Good opportunity to say a warm 
Thank You.

-- 
Thomas

Post a reply to this message

Il 14/01/2023 01:50, Samuel B. ha scritto:

Thank you, Sam! A different approach can be a stimulus to revisit old 
projects (and I have many, lost in time...).
Btw, a nice and relaxing test image.

Paolo

Post a reply to this message