Page 15 of 17

Min. Epsilon: Metropolis, GPU Path vs CPU Path

Posted: Sat Sep 22, 2018 2:10 pm
by FarbigeWelt
Modelling a new ashtray I've noticed some glitches in view port render which have been removable increasing min. epsilon by factor 100.
Now, testing Metropolis regarding sun light and glass windows I've noticed a glitch, dark line below second window from left, in the final render when using openCL. The first reliable min. epsilon for GPU is 0.0001. This glitch cannot be observed using CPU at default min. epsilon 0.000001.
2018-09-22 GPU Path min. epsilon 0.000001.png
2018-09-22 GPU Path min. epsilon 0.000001
2018-09-22 GPU Path min. epsilon 0.0001.png
2018-09-22 GPU Path min. epsilon 0.0001
2018-09-22 CPU Path min. epsilon 0.000001 (2).png
2018-09-22 CPU Path min. epsilon 0.000001 (2)
Does CPU and GPU render with different precision? If yes, is higher precision on GPU possible?
If yes, what about implementing a checkmark to set precision for GPU?

Wooden Room, improved lighting

Posted: Mon Sep 24, 2018 9:36 am
by FarbigeWelt
CPU BiDir, Metropolis, 1xHD
Improved lighting with new lamps. Different ashtray model. Rendering with 2xHD is still ongoing (540 Samples after 7 h). Noise reduced in this picture with photoshop.
First time overall lighting looks like I expect it. Table lamps look now as they should based on rough metal.

Re: Test scences GPU Path vs CPU BiDir

Posted: Mon Sep 24, 2018 9:55 am
by B.Y.O.B.
FarbigeWelt wrote:
Sat Sep 22, 2018 2:10 pm
Does CPU and GPU render with different precision? If yes, is higher precision on GPU possible?
Both render with 4-Byte floating point numbers.
Maybe Dade can tell you more about why GPUs often show artifacts earlier and require higher min. epsilons.

Wooden Room, CPU BiDir, Metropolis, 1191 Samples, 2xHD

Posted: Mon Sep 24, 2018 8:26 pm
by FarbigeWelt
Wooden Room, CPU BiDir, Metropolis, 1191 Samples 22h26m, 2xHD.jpg
Wooden Room, CPU BiDir, Metropolis, 1191 Samples 22h26m, 2xHD
After almost one day render time, I saved image as openEXR 32 bit float and converted it to 16 bit in Photoshop, also removed some bright single pixel spots with remove scratches filter and finally exported as JPG compressed 67% to current file size. There is still some noise there but I am pleased with the overall result. Converting from 32 bit to 16 bit required some gamma and brightness correction.

The test picture in the back shows a jump in the cyan blend and hard to distinguish RGB tiles close to saturation. It is actually close to what the orignal picture shows on my screen. (Seems like the test picture could use a little update. Actually I created it to test my printer's setting after an unfortune driver update. But it also helps to adjust render settings or post process workflow.)

The wood texture is procedural, all the same basis with different offsets and factors.
Glasses are default with IOR of crown glass pure.
Metals are default with different roughness for table lamps, picture frame, ceiling spots.
The spots are made from two thin, joined mesh objects; an inner with matte white 1,1,1 an outer with metal, both based on interior clear IOR 1.6 volume. The same volume is used for wood and glass too.

(First, I defined two different materials for the same objects but then I noticed light passes through these objects, especially with Metropolis but also with Sobol. Inner layer was matte and outer was metal both pointing to the same interior volume, clear with IOR 1.6. My current conclusion is: Avoid this kind of object definition.)

As mentioned in earlier posts, correct lightning of a scene is essential. The panels in front of the ceiling lamps avoid direct, dazellling light. The panels' shape is the fourth generation, a good starting point for a next generation. Ceiling and walls are used for indirect light, knowing wood is not really appropriate, but this is a reason why wooden rooms look dark and rather small and are difficult to be illuminated in reality. (Hard to tell, I prefer bright rooms, why some people like to live in wooden rooms apart from their comfortable climate. But for this virtual Alp Bar wood fits well.)

Currently the point lamps' settings are: gain 1.0, power 100, efficacy 100.

Re: Test scences GPU Path vs CPU BiDir

Posted: Fri Sep 28, 2018 8:46 am
by FarbigeWelt
Dozens of different settings test wete required to get a reasonable result using parh, metropolis. Meanwhile I created a new test picture to get useful information for improving. The rendering is still ongoing to see if current setting leads to acceptable noise. Parameters I had to adjust: clamping and metropolis settings. What I observed with clamping was the grey shades 25, 50, 75, 100% looked all the same, color gradient from 0 to 100 R, G or B despite of black also. I used natural numbers only in the input field for clamping. I will test what happens if I enter e.g. 200.01 because I have the impression path does not use any detailed color shades in my renders neither cpu nor gpu.

Images will follow this evening.

Found an already quite optimized gpu path render. The image is cropped, 3500 samples, clamped. Lok at the blue gradient there seems to be one blue shade only, darker tiles are dithrered and tiles have almost the same shade.
Test image in wooden room scene, cropped

GPU Path Metropolis, 4000 Samples

Posted: Fri Sep 28, 2018 10:03 pm
by FarbigeWelt
Render finished with 4000 Samples for 2xHD after 6 h. Image is very clean, maybe a bit too dark. I tried to adjust brightness and gamma to get a brighter version but then color gradients did not fit anymore. Metropolis with CPU BiDir and now GPU Path works well for this scene. After I found acceptable settings for sobol and path tile I will improve some details and upload the scene.
Inside Alp Bar, openCL Path, Metropolis.png
Inside Alp Bar, 4000 Samples
Inside Alp Bar, GPU Path, Metropolis, 53 Samples.png
Same scene and settings as above but with 53 samples

openCL Direct Light Caching with 9000 Lamps

Posted: Sat Sep 29, 2018 9:36 am
by FarbigeWelt
The openCL support of direct light caching and its speed are genious! Thanks a lot to all.
GPU Path Sobol, DLCS, 1017 Samples.png
GPU Path Sobol, 1017 Samples
Test image converges in short time, noise level in dark areas require up to 1000 samples with sobol.
In this scene color gradients and grey shades are rendered nicely with openCL.
Sine curved LED
Sine curved LED lamp
(376.25 KiB) Downloaded 71 times
Feel free to use the attached scene for any improvement.

Alp Bar Blender Scene

Posted: Sun Sep 30, 2018 11:54 pm
by FarbigeWelt
Alp Bar GPU Path Sobol, 500 Samples denoised, 2xHD, 8.png
Alp Bar GPU Path Sobol, 500 Samples denoised, 2xHD, 8
The Alp Bar Blender file can be downloaded here: ...

There are cameras for inside and outside view.
CPU BiDir requires tonemap gain of 0.02 and clamping of 2050.
GPU Path requires tonemap gain 0.125 and clamping between 205 and 405.
Denoiser can be used for Sobol with pretty results (CPU 100 to 200 samples, GPU 250 to 500 samples).

Comparision of raw and denoised closeups from 2xHD resp. 1080x1080 images.
Alp Bar GPU Path Sobol, 500 Samples comparison, 2xHD.png
Alp Bar GPU Path Sobol, 500 Samples comparison, 2xHD
Indirect Light, GPU Path Sobol, Denoiser Test, 1000 Samples.png
Indirect Light, GPU Path Sobol, Denoiser Test, 1000 Samples
Conclusion: Denoiser preserves details astonishingly accurate.

Joined Meshes and Particle System with strange effects

Posted: Wed Oct 03, 2018 5:40 pm
by FarbigeWelt
The approach of setting up this scene leads at the edge of my PC. Have to change a lot. However, particle system should also be correct if view port visibility is disabled. Actually all leaves are based on same material setting despite the color. As you can see two of the trees look rather special. The red brown tree is the reference. Its leaves are duplicates created with sapling tree gen. The leaves of the other trees are real made duplicates, made single user (object and data), joined meshes, randomly selected vertexes, selected linked and finally seperated selected. The separated leaf meshes are parented then to its tree object again. Blender's memory usage increased to swapping 34 GB and blender file takes about 2 GB. The two leaves should be glossy translucent but are not rendered accordingly. The particle system should show grass but one can see blocks. VRAM used during render was 3.5 GB with denoiser enabled, GPU path. So much to this approach. I start investigating the two render issues with a handier scene soon.
Three Trees strange effects.png
Three trees, two with strange material effect, blocks instead of grass

Glossy and strange Glossy Transluscent

Posted: Wed Oct 03, 2018 8:18 pm
by FarbigeWelt
Despite of transmission color same settings for both glossy transluscent and glossy. In my opinion there is something wrong with specularity in glossy transluscent material.
Glossy and Glossy Transluscent.png
Glossy and Glossy Transluscent
Test Joined Mesh
Test Joined Mesh simple
(96.94 KiB) Downloaded 68 times