Cycles Development Updates

And that is the issue. Not everyone is using cycles to make animations. Not every feature works well for both stills and animation.

edit: at least that is my issue, Theory seems to be getting good mileage out of scrambling distance for animations.

1 Like

In all of those cases, the bias created (if used with the right settings) will barely be noticeable compared to a full 30 percent difference in brightness (which would really make the Cycles render the odd one out if compared to similar results from other engines).

The thing about what Lukas mentioned about brightness differences could really mess with you if you were doing stills and wanted the result to look correct (because you couldn’t really tell if the brightness is overdone or not). I also find it frustrating that people would actually want features with major flaws because they have a workaround in mind (to heck with making sure that features just work when they debut in an official Blender release).

But if there is no equilibrium in the release of features to use for animations.
Maybe, nobody will use it at all for that.

Ace, have you used a build with scrambling distance?

I posted a windows build a couple days ago. Play with it a bit. It certainly doesn’t fix everything (like clamping, filter glossy, and denoising don’t fix everything) but it can really help in some scenes(like clamping, filter glossy, and denoising help in some scenes), and it especially helps in the viewport. even as just a viewport option, it would be very helpful.

3 Likes

even Ai denoiser’s main goal is to accelerate the viewport, thus better the UI flow
as times change, goals change, intentions change, philosophies change - only truth stays

reminds me of…

“I disapprove of what you say, but I will defend to the death your right to say it.”

but here it seems the opposite prevails

when the leading entity accepts it, then the weak are ‘free’ to follow

I was under the impression Blender is meant for Animation.
:slight_smile:
keep it up

A long time ago, setting it to a low value creates a lot of initial bias in scenes with multiple light sources and it can take ages to clear all of that up, especially if you want a correct looking result.

It’s too bad really, it would’ve been good for the denoiser if only that was solvable.

The lower you set it, the more bias it introduces. if you set it to something less drastic, it can have some really good results

Here is a test that I put together, using one of your favorite bugaboos: Caustics!
(click for comparison)


This is a groundplane under a glass suzanne, 16 bounce GI. 1000 samples.

Changing the scrambling distance to .1 renders ~70% faster, with less noise (10,000 sample render provided for comparison)

This is just one case, and absolutely there are cases where it isn’t the right tool for the job, but the same is true for clamping, filter glossy, and denoising.

6 Likes

The difference though is that clamping, filter glossy, and denoising can be close to being useful in a universal sense. What it means is that no matter what scene you have, having these features enabled (but with values not too drastic) can only help.

For reference, I use ‘filter glossy’ with a value of 0.2 and clamping at 50 and 100 respectively. With branched patch tracing, this can result in nice images with little bias. Denoising is a little trickier, as that feature still has quality issues in areas like highlights and small details (which any realistic scene will have). Hopefully, we can take ideas from Luxcore’s denoiser and resolve them.

Scrambling Distance at 0.1 might be doable with your example because there’s only one light source, the quality issues start when the sampler has to pick from multiple sources (they can go away with higher values like 0.75, but they otherwise eat a lot of the supposed time savings).

Thank you!

Oh wow! That’s awesome! I had no idea the Threadrippers were that fast. Thanks for posting these guys

While that was certainly a test that plays into the strengths of scrambling distance, rendering 70% faster is not a supposed time saving. That is drastic. other scenes are less drastic, some have no speedup at all, but overall, it still nets a noteworthy (not simply supposed) speedup:

17% faster:

23% faster:

5% faster:

no change:



It’s certainly not a magic bullet, but it is beneficial. I can understand not wanting to confuse the user base, but that is exactly what the experimental option should be used for.

4 Likes

I tested scrambling as well. It uses 2 area lamps in windows, a sun lamp and a HDR. Lower scrambling values produce artifacts in lighting, noise level seems about the same, but times are quicker using scrambling. I find values around 0.3 to be good for this scene (there are small differences in lighting but without reference it wouldn’t be noticeable).
I’m more interested in the GSOC multiple lamp sampling enhancements which looked more generally usable.
Click the image for full scale (9MB).

5 Likes

First I want to say that I trust the developers 100 percent what features to include and when. I’m very thankful for providing us with this great software!

I’m also testing the patch with Scrambling Distance and from my current perspective it seems to be a really valuable tool to bring down rendertime and/or images with less noise. But I have to test it some more.

Here is a low/medium complexity scene with 200 samples BPT (no Denoising). This shot has 5 main light sources, 97 small mesh lights and Volume Effects all over the place (click for Full HD):

Scrambling Distance 1 (original Cycles Sampling), Rendertime 13:04.69

Scrambling Distance 0.1, Rendertime 10:06.88

Scrambling Distance 0.002, Rendertime 08:37.19

Even the rather extreme value of SD 0.002 would be almost acceptable in this case. For final rendering I would go 50-100 samples higher and the remaining artifacts and noise would be smoothed by the Denoiser, which seems to work very well with Scrambling Distance. The temporal stability seems to be good and is way better than I thought, but I will test this more to be sure.

In my other tests I found that Scrambling REALLY improves Volume, SSS and transparent materials, which are main problems for the Denoiser. The Scrambling artifacts are also getting smoothed by the Denoiser. They work well together.

If this patch gets included I would also propose to put it under experimental, maybe in the simplify panel to hide it some more.
If it doesn’t get included, I would be fine if Lukas updates the patch once every 1 or 2 months to keep it working. Then we can build it ourselfes.

5 Likes

Some of the more complex examples indeed show that very low values (such as 0.1) appear to cause certain areas of the image to become a little darker (not much, but enough to be noticeable).

However, higher values of 0.3 or above cuts the bias dramatically, but then you end up cutting a decent chunk off the time saved as well. That was what I saying about whether it would really save that much time (because you would need a higher number of samples to clean up said bias. I have also noticed that motion blur and shadows in a couple of examples seem to start showing “steps” or slight discontinuities in what should be a smooth falloff.

Now the bit about the denoiser should be true because of the way the sampler is now focused a bit more on being even. I do wonder if a more even or grid-based sampling feature could be coupled with the denoising feature so as to make it easier to get good results (could be based on the scramble distance patch, but better quality).

Thanks for the build !
what i love the most about SC is that the view-port becomes very smooth with sss or volumetric, sc at 0.3 with more sample is better than less sample but very laggy viewport

That first shot in high castle of nyc has a light in every single windows theres over 500 lights there.

1 Like

What are you rendering on?

It’s an older hexacore i7-5930k, 32 GB ram, 750ti 2gb, and a Quadro k4200.

I definitely strayed from defaults on the benchmark scenes, I didn’t want to spend all day benchmarking. Most of the scenes were rendered at 50% resolution, and reduced sample count.

But did you do cpu or gpu or both

Bmw, classroom and Barcelona were rendered on CPU/GPU, the rest were on CPU. I’ll do some more involved testing when I get a chance