Test scences GPU Path vs CPU BiDir

Post your tests, experiments and unfinished renderings here.
User avatar
FarbigeWelt
Posts: 362
Joined: Sun Jul 01, 2018 12:07 pm
Location: Switzerland
Contact:

Two Beers and a Mirror

Post by FarbigeWelt » Thu Jul 26, 2018 7:52 pm

Foam is built with even more bubbles. Info notes 500 MTris, 200 kS/s, 90 Samples.
Two Beers 90 S CPU BiDir Sobol Depth 50.jpg
Two Beers 90 S CPU BiDir Sobol Depth 50 1080
EDIT 1
Now, the same picture but rendered with 2160x2160 and reduced to 1080x1080. Spot the difference.
Two Beers 90 S CPU BiDir Sobol Depth 50 2160 reduced to 1080.jpg
Two Beers 90 S CPU BiDir Sobol Depth 50 2160 reduced to 1080
Okay, both pictures are filtered with remove scratches, 1 pixel size.

Next step is 4320x4320.

EDIT 2
80 Samples, 200 kS/s, 2 hours
Two Beers 80 S CPU BiDir Sobol Depth 50 4320 reduced to 1080.jpg
Two Beers 80 S CPU BiDir Sobol Depth 50 4320 reduced to 1080
I wonder if 16x80 Samples for a 1080 image will be comparable to the one above.
Microsoft Windows 10 Professional, intel i7 4770K, 32 GB, AMD R9 290x 4 GB, AMD R9 390x 8 GB
Instagrammer, please join #luxcorerender for your renderings.

User avatar
FarbigeWelt
Posts: 362
Joined: Sun Jul 01, 2018 12:07 pm
Location: Switzerland
Contact:

Multi Samples vs Reduced HiRes

Post by FarbigeWelt » Fri Jul 27, 2018 5:45 am

Two Beers 1280 S CPU BiDir Sobol Depth 50 1080.jpg
Two Beers 1280 S CPU BiDir Sobol Depth 50 1080
Well, obviously the picture from the last post shows much more details than the one above. I've assumed such a result based on several observations during fireflies fighting sessesion. The 4x resolution rendering (16x pixels) is about then percent slower than 1x resolution.

Conlusions: One gets more details with higher resolution. This is an easy fact. Getting a more detailed image if reduced to smaller resolution is less obvious. One may expect that sampling with small x,y variations and adding color values within number’s bit resolution should average out noise. I guess there is some room for improvements in the render algorithm.
Microsoft Windows 10 Professional, intel i7 4770K, 32 GB, AMD R9 290x 4 GB, AMD R9 390x 8 GB
Instagrammer, please join #luxcorerender for your renderings.

User avatar
B.Y.O.B.
Developer
Posts: 1825
Joined: Mon Dec 04, 2017 10:08 pm
Location: Germany
Contact:

Re: Test scences GPU Path vs CPU BiDir

Post by B.Y.O.B. » Fri Jul 27, 2018 7:46 am

Note that with supersampling you reduce the filter radius (indirectly) when you downscale.
You could try to use a smaller filter radius in the normal-resolution image, this will make the image appear sharper.

Did you save the high-resolution image as 32 bit EXR, and did the program you used for the downscaling work in 32 bit depth?
If not, you are also performing an implicit clamping operation (which would explain the darker image and removed fireflies).
Support LuxCoreRender project with salts and bounties

User avatar
FarbigeWelt
Posts: 362
Joined: Sun Jul 01, 2018 12:07 pm
Location: Switzerland
Contact:

Re: Test scences GPU Path vs CPU BiDir

Post by FarbigeWelt » Fri Jul 27, 2018 5:12 pm

B.Y.O.B. wrote:
Fri Jul 27, 2018 7:46 am
Note that with supersampling you reduce the filter radius (indirectly) when you downscale.
You could try to use a smaller filter radius in the normal-resolution image, this will make the image appear sharper.
Usually I render without any filter.
B.Y.O.B. wrote:
Fri Jul 27, 2018 7:46 am
Did you save the high-resolution image as 32 bit EXR, and did the program you used for the downscaling work in 32 bit depth?
Program can work with 32 bit files but I saved standard png. I can render a 32 bit EXR for comparision.
B.Y.O.B. wrote:
Fri Jul 27, 2018 7:46 am
If not, you are also performing an implicit clamping operation (which would explain the darker image and removed fireflies).
The darker image anoyed me. You support my idea to use another bit depth and floating points instead.
Photoshop has a very nice filter for removing scratches. I appöied it on all pictures above.

What I observed the higher the resolution the smaller the firefly because it is often only one superbright or wrong colored pixel. Downsizing with averiging or smarter algorithms smooth the fireflies but hardly remove them.

What I currentky cannot understand is the increas of details. Look at the foam or the water drops on the upper part of the beer glasses.

(I still don‘t like the look of the water on the glass. Blender fluids simulation is quite limited.)
Microsoft Windows 10 Professional, intel i7 4770K, 32 GB, AMD R9 290x 4 GB, AMD R9 390x 8 GB
Instagrammer, please join #luxcorerender for your renderings.

User avatar
B.Y.O.B.
Developer
Posts: 1825
Joined: Mon Dec 04, 2017 10:08 pm
Location: Germany
Contact:

Re: Test scences GPU Path vs CPU BiDir

Post by B.Y.O.B. » Fri Jul 27, 2018 5:32 pm

FarbigeWelt wrote:
Fri Jul 27, 2018 5:12 pm
(I still don‘t like the look of the water on the glass. Blender fluids simulation is quite limited.)
For my beer glass, I modeled a few different drop shapes, then distributed them on the outer glass surface with a particle system.
Support LuxCoreRender project with salts and bounties

User avatar
FarbigeWelt
Posts: 362
Joined: Sun Jul 01, 2018 12:07 pm
Location: Switzerland
Contact:

Re: Test scences GPU Path vs CPU BiDir

Post by FarbigeWelt » Fri Jul 27, 2018 8:50 pm

B.Y.O.B. wrote:
Fri Jul 27, 2018 5:32 pm
For my beer glass, I modeled a few different drop shapes, then distributed them on the outer glass surface with a particle system.
Thank you for this input. I already thought about to buy the new blender add on FLIPS. I will try the demo version first to look if it is faster than built in fluidics. But modelling and using particle system sounds more appropriate.
Microsoft Windows 10 Professional, intel i7 4770K, 32 GB, AMD R9 290x 4 GB, AMD R9 390x 8 GB
Instagrammer, please join #luxcorerender for your renderings.

User avatar
FarbigeWelt
Posts: 362
Joined: Sun Jul 01, 2018 12:07 pm
Location: Switzerland
Contact:

Closer look at over sampling

Post by FarbigeWelt » Fri Jul 27, 2018 9:31 pm

Pictures are rendered with animate and saved with openEXR Float (full). Then I imported the picture in Photoshop copied the right version to a 16 bit image with the doubled horizontal resolution. After that I converted the image to 16 bit because the filter "remove scratches and dust" is not available in 32 bit. The filter was applied with radius 1 and threshold 100 eliminating firelies but also reducing sharpness. Threshold could be increased for higher resoluting pictures but for comparablity I've choosen the same setting for all images. Finaly I saved the files as png, all with the resolution 256x512. For the comparision I've copied them into the 512x1024 image here:
1024 S90 512 S90 256 S1440 256 S90.png
1024 S90 512 S90 256 S1440 256 S90
Still the pictures do not have the same brightness. I'm not very familiar with Photoshop maybe there is way to apply adjustments to one layer only.

However, it seemed first to me that pictures show more details from rendering at higher resolutions and downsizing, than increasing number of samples at lower resolutions. But there is an advantage of more samples I did not recognize Look at the third row, the image has more volume like more layers of applied colors. Whereas the first row's pictures look kind of hollow.
Microsoft Windows 10 Professional, intel i7 4770K, 32 GB, AMD R9 290x 4 GB, AMD R9 390x 8 GB
Instagrammer, please join #luxcorerender for your renderings.

User avatar
FarbigeWelt
Posts: 362
Joined: Sun Jul 01, 2018 12:07 pm
Location: Switzerland
Contact:

Beers again

Post by FarbigeWelt » Sun Jul 29, 2018 12:03 pm

2 Beers HD 400 S 200 kSps Depth 50.png
2 Beers HD 400 S 200 kSps Depth 50
Info Blender: Faces: 129 k, Tris: 260 k, Mem: 1455 M
Info LuxCoreRender: BiDir CPU + Sobol, 776 M Tris
Output: OpenEXR RGBA Float (full)
Changed to 16 bit in Photoshop with some adjustment in the curve diagram.
Removed part of noise with filter, enhanced edges with filter and smoothed with small radius.
Exported as 8 bit png.

Part of these adjustments have been done because foam was heavy overexposed. Rendering is ongoing with small adjustments like i.e. foam's specularity.

EDIT:
Depth: 60, GPU Path: 1 MS/s 256 Samples, but with very different result. Interesting to see how many reflections between the mirrors appear here. Blue color is gone. One can see some rest of the blue color at the bottom up reflector of the light.
2 Beers HD 256 S 1 M Sps Depth 50 GPU Path post processed.png
2 Beers HD 256 S 1 M Sps Depth 60 GPU Path post processed
Microsoft Windows 10 Professional, intel i7 4770K, 32 GB, AMD R9 290x 4 GB, AMD R9 390x 8 GB
Instagrammer, please join #luxcorerender for your renderings.

User avatar
FarbigeWelt
Posts: 362
Joined: Sun Jul 01, 2018 12:07 pm
Location: Switzerland
Contact:

Glossy with snow, Matte boring

Post by FarbigeWelt » Mon Jul 30, 2018 7:37 pm

2 Beers HD 2048 S 200 kSps 13h40 Depth 60 CPU BiDir.png
2 Beers HD 2048 S 200 kSps 13h40 Depth 60 CPU BiDir
2 Beers HD 2048 S 260kSps 2h50 Depth 60 CPU BiDir.jpg
2 Beers HD 2048 S 260 kSps 2h50 Depth 60 CPU BiDir
Both files have been save as openEXR and post processed manually, optimized tone and removed some noise.
Although it is very noisy, I prefer the impression of the first picture. Not expected is the big difference oof total rendering time considering the small difference in kSps. In both cases 2048 samples are not enough to resolve image in the deepths of reflections. Looking at these gaps I guess the image requires 10 kSamples for the matte material and countless samples for glossy. It may take some years until I own an adaquate computer to render a final version of the scence in a reasonable time. Again, I struggled with glossy material. A topic I will look into in the next few posts.
Microsoft Windows 10 Professional, intel i7 4770K, 32 GB, AMD R9 290x 4 GB, AMD R9 390x 8 GB
Instagrammer, please join #luxcorerender for your renderings.

User avatar
FarbigeWelt
Posts: 362
Joined: Sun Jul 01, 2018 12:07 pm
Location: Switzerland
Contact:

A new approach: GPU Tile Path

Post by FarbigeWelt » Mon Jul 30, 2018 10:35 pm

There is probably room left for further improvements in a reasonable time. :D
GPU Tile Path T256 AA8 1.9MSps 10 min 2HD.jpg
GPU Tile Path T256 AA8 1.9MSps 10 min 2HD
Microsoft Windows 10 Professional, intel i7 4770K, 32 GB, AMD R9 290x 4 GB, AMD R9 390x 8 GB
Instagrammer, please join #luxcorerender for your renderings.

Post Reply