Another test, this time with surface normals, extra material properties, and
photons. It took 44 seconds to render. The metallic objects and the glass
spheres got a tiny granite normal; the plane got quilted. The result isn't too
shabby, but it's missing some details here and there (most notably, the
horizontal lines on the plane normal are almost gone).

I do think this is workable solution for getting high-quality renders using low
(and fast) settings, if one doesn't mind not being a POV-purist :)

Post a reply to this message

Attachments:
Download 'denoise-pov-ray2.jpg' (134 KB)

Preview of image 'denoise-pov-ray2.jpg'

"Samuel B." <stb### [at] hotmailcom> 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.
>
> The scene was rendered in UberPOV using no_cache radiosity and a count of 1, and
> 8 blur_samples for the camera. No AA. (It rendered in only 14 seconds. The
> denoising in Blender took a little over a second.)
>
> Works nicely, eh? I don't know yet if it will handle media, transparent objects
> and other such things, but this initial result is promising!
>
> Sam

Is this based on some kind of resampled importance sampling, with spatiotemporal
resuse - like "ReSTIR"?

I was just watching a presentation on this kind of thing not that long ago, and
looking over some of the papers detailing how it all works.

https://www.youtube.com/watch?v=gsZiJeaMO48


Good to see you around again, and still doing very interesting experiments   :)


- Bill

Post a reply to this message

"Bald Eagle" <cre### [at] netscapenet> wrote:
> Is this based on some kind of resampled importance sampling, with spatiotemporal
> resuse - like "ReSTIR"?
>
> I was just watching a presentation on this kind of thing not that long ago, and
> looking over some of the papers detailing how it all works.
>
> https://www.youtube.com/watch?v=gsZiJeaMO48
>
>
> Good to see you around again, and still doing very interesting experiments   :)
>
>
> - Bill

Hi Bill,

Thanks, it's good to see you, too!

Hopefully I'll get more stuff going soon. I tend to go through long stretches of
having no creativity... or rather, I often lack the energy needed to make
things. One problem is I've always tried to do everything except mesh modeling
in POV-Ray, but that resulted in me having to do crazy things just to get
reasonable render times. With this denoising, I can hopefully stop relying on
many of those crazy tricks.

I don't know the details of how Blender denoises images, but I think it uses
Intel's Open Image Denoise, which is based on machine learning of some kind.

Sam

Post a reply to this message

"Samuel B." <stb### [at] hotmailcom> 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.
>

That's a pretty amazing trick and result. Nicely done. (I still haven't yet
dabbled with Blender, but it's on my 'bucket list'. Along with 100 other
things...)

Aside from the slightly off-color whites in this first denoised image (which you
solved), I see a slight greenish halo around the upper-left emissive sphere. I
guess that's to be expected from the heavy original focal blur and low
blur_samples count. But the result actually looks to my eyes like what might
happen with a real camera lens, using film. So I like it!

And your later image with the normals et al is really nice.

Post a reply to this message

Cave test with height fields. It took a little over 5 & 1/2 minutes to render
using lowish settings and media.

This time I made use of normal and albedo passes in order to preserve the
details. (And as it turns out, the final result is only good when using the same
camera aperture for every pass.) In addition to the denoising, I: corrected the
gamma; added glare; and added a bit of film grain. The scene itself could use
some work :D

Post a reply to this message

Attachments:
Download 'denoise-complex-pov-ray2-5m_32s.jpg' (107 KB)

Preview of image 'denoise-complex-pov-ray2-5m_32s.jpg'

"Samuel B." <stb### [at] hotmailcom> wrote:
> Cave test

Here's an image showing the three passes that came out of POV-Ray:
  1) normals
  2) albedo
  3) full render

I used some #if and #switch statements throughout the code which allowed me to
just run a 3-frame animation to get all the passes.

If you compare the rock's face (on the left) to the previous image I posted, you
can get an idea of how much the normal and albedo data helps the denoiser make
sense of lots of noise.

Post a reply to this message

Attachments:
Download 'denoise-complex-pov-ray2-passes2.jpg' (155 KB)

Preview of image 'denoise-complex-pov-ray2-passes2.jpg'

"Kenneth" <kdw### [at] gmailcom> wrote:
> That's a pretty amazing trick and result. Nicely done. (I still haven't yet
> dabbled with Blender, but it's on my 'bucket list'. Along with 100 other
> things...)
>
> Aside from the slightly off-color whites in this first denoised image (which you
> solved), I see a slight greenish halo around the upper-left emissive sphere. I
> guess that's to be expected from the heavy original focal blur and low
> blur_samples count. But the result actually looks to my eyes like what might
> happen with a real camera lens, using film. So I like it!
>
> And your later image with the normals et al is really nice.

Hi Kenneth,

Thanks!

The denoiser is doing all the heavy lifting; everything else is pretty easy when
using just one pass.

Learning how to operate Blender might be worth it, even if only for its modeling
capabilities. But it's admittedly not an easy program to get into, mostly due to
the daunting interface. But it all starts to make sense and eventually shows
itself to be a good design choice once you learn the basics, imo :)

Sam

Post a reply to this message

Le 2023-01-13 à 17:33, Samuel B. a écrit :
> 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.
> 
> The scene was rendered in UberPOV using no_cache radiosity and a count of 1, and
> 8 blur_samples for the camera. No AA. (It rendered in only 14 seconds. The
> denoising in Blender took a little over a second.)
> 
> Works nicely, eh? I don't know yet if it will handle media, transparent objects
> and other such things, but this initial result is promising!
> 
> Sam

Look like a simple blur operation using a 3x3 block size.

Post a reply to this message

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'