Cycles Development Updates

(Flavio Della Tommasa) #888

Do you plan to implement the Lamp Sun in order to have the lighting according to its inclination [such as in Vray] so you can recreate sunrise / sunset?
Making an animation in blender for the Solar study is not the best of simplicity…

(SterlingRoth) #889

There is a sun angles addon out there somewhere that works well for this

(Flavio Della Tommasa) #890

Hi Sterling!
Do you mean the Sun Position addon?

(J_the_Ninja) #891

It doesn’t affect the color of the sun lamp itself though. Also doesn’t have a visible solar disk when viewed directly.

(CarlG) #892

I think the color of the sun is about 5500k, no matter what’s its altitude is. As for reflection, the sun size is about 0.005 for the disk itself, but then you want to include the suns corona, atmospheric conditions, and even overcast - these kinds of gradients are probably not that easy to make.

It’s the atmosphere that causes a color change. For mars and moon scenes things would be different. So it’s more the sky texture that would have to change.

But even better, a proper physical sun&sky “system” using absolute scale values for brightness. The sun would be basically an indicator and main lightsource direction, whereas the sky texture component should react properly to the sun orientation (incl brightness which is also “wrong” wrt absolute values). The obvious caveat here is when you need a parallel light which is not a sun or a scifi scene with more than one sun.

I’m not sure what is the best approach here. If it ends up a “big texture” it might have a sampling cost compared to the current limited system (which is why I prefer sun&sky, albeit crappy, over HDRIs, to do the lighting for my interiors in the few cases I actually need sun&sky or just sky).

(Piotr Adamowicz) #893

5500k is just a ballpark figure for what gets through our blue-scattering atmosphere on average afaik. The light is much bluer in space, and much redder at sunset when the light needs to get through a thicker layer of air.

(CarlG) #894

I’m not sure about much bluer? Acc to this it’s 5778K in space. I just use 5500k because it is an easy starting point to remember :smiley: So it’s even slightly more towards red than standard white of 6500K, but it’s white enough so we just say it’s white. Sun is a class G which is yellow, a more white star would be class F. But if you look at the spectrum, the sun is actually green because it peaks at around 500nm (edit). I admit, I never heard of this before… :stuck_out_tongue:

But I’ll call the sun yellow, just because of its classification as such.

(moony) #895

Noticed something weird with Cycles today.

I got an Oculus Go for Christmas and wanted to experiment with 360 degree stereoscopic VR rendering. I modified a BMPS scene to output a 360 degree stereoscopic JPEG and rendered a 2048x1024 pixel image using a tile size of 32x32, 50 samples and denoising turned on. I was using one of the daily 2.79 builds on the 2.8 branch that has combined GPU/CPU rendering activated (27th Sep 2018 build).

I rendered the scene normally, allowing the render result to build up on screen and got a render time of 7 minutes 32 seconds across both stereoscopic frames.

I then re-rendered the scene, but just after starting the render - switched to render slot 2. This just showed a grey screen, but showed the render tiles outlined in orange as the render built up in the background.

The weird thing is however, the render proceeded much faster. I got a total render time for the scene of 1 minute 13 seconds for the combined stereoscopic image (around 6x faster). I switched back to slot one after the render had finished and as far as I can tell - the two renders are identical in resolution, noise etc.

I have uploaded my test file in case anybody wants to confirm my finding.

It’s almost as if updating the tile on the on screen image causes a massive performance drop. By switching to slot 2, the on screen updates don’t happen and the render speed is significantly enhanced.

edit: after a bit more experimentation - it appears that it may be the stereo view on the render output that is causing the slow down. If I click the “display the image in stero 3d” button (the one that looks like a pair of anaglyph glasses) - so that the single image in displayed, not the anaglyph stereo one - the render speeds up significantly.

(moony) #896

Incidentally - I did try rendering the scene above using 2.8 beta, but it crashes after the first stereoscopic image pass.

Looks like 2.8 beta might have a bug with stereoscopic panoramic rendering.

(Lsscpp) #897

Rendering without seeing the image buildup is actually faster, not that much though

(pixelgrip) #898

and saves memory too,if you switch from render preview to solid for example,before hit F12

(Bernardo) #899

I am not sure about 2.8 but in 2.79 when you render a stereo image there is a small button in the image editor toolbar that is active by default and when on, makes the rendering times skyrocket. I think the only thing that little button is doing is showing the rendered image as stereo 3d, but I think there’s something wrong with it. The image does show as stereo, but I don’t see why it should have such a ridiculous impact on render times. Anyway… try to look for that little button and turn it off.

(moony) #900

Yep - I already did that in my edit at the bottom of the original post.

(Bernardo) #901

sorry, I missed that :slight_smile:

(Zsolt) #902

Yes, I reported this on the bug tracker about a year ago. It was confirmed as a bug, but hasn’t been fixed.
Using larger resolutions (4K+) makes an even larger difference. So I just disable the anaglyph glasses button. The default setting for this should be OFF.

EDIT: found it:

(Ace Dragon) #903

Lukas Stockner just pushed an experimental build with his animation denoising patch.

Overview just in case.

(brothermechanic) #904

hello all
i did build with embree patch
this is result
without embree

with embree

What do you say?


Embree should give better speed and memory usage at complex scenes with motion blur.
This simple scene is probably not the best test.

(smilebags) #906

Yeah, Embree is all about optimising rendering performance and memory usage in large scenes and motion blur. This is neither a large scene nor contains motion blur so you’re unlikely to see any significant differences.

(Thesonofhendrix) #907

Question… Will or can the tensor cores on an RTX or other tensor device be used to calculate cycles path tracing? If so is it faster than cuda cores and would it be worth just buying a TPU?