What's the magic behind One-sample High-D ?

[Bzuco]

I was trying new psycho sampler and I was surprised how fast it can clean up the noise. So I tested all samplers and choice  One-sample High-D, because it it clean up noise much faster than the default. I dont see any visual problems in interior and exterior daylight renders yet. Are there some disadvantages I dont know about?
And what ist the magic behind this sampler, why it clean noise much faster than default one?
What can be controlled in rendering process with random samplers?
Could we have another random sampler for test renders which will produces render with less correct GI(or fake GI) but almost noise free image after few passes ...something like vray with extremely low GI settings.

Too many questions :)) , but I hope will get answers on all of them. Thanks.

[Ondra]

can you post some example comparison? AFAIK the problem with some of the non-default samplers is that they just stop converging after a while in some scenes

[Bzuco]

In interior scene it is visible on floor under window, wall near gray teapot between torus, and floor under direct sunlight. It is visible more in direct light parts than in indirect.

20pass, 30s


20pass, 30s


20pass, 12s


20pass, 12s

[Bzuco]

comparison in bigger ext. scene
both 16pass, 30s

[maru]

That's really odd, I would never expect such results.
What are your other render settings, did you change anything from defaults?
What kind of enviro lighting are you using in your scenes?

There is a lot more noise visible in the direct light cast through the window with the torus knot example. And this is something we would expect to be much cleaner much faster with the new sampler.

[Ondra]

it is refreshing to see a report with surprising speedup instead of surprising slowdown ;)

last time i tested that sampler it was generally working bad. I guess it is time to re-evaluate and maybe finish it up (it is still AFAIK pretty unoptimized)

[Bzuco]

My render settings for exteriors:
total passes: 192
GI/AA balance: 3
light samples mult. 0.5
max ray depth: 6
Adaptivity: off
portal fraction: 1.0
Embree precise and quality both off
denoising: off
exposure: -4
- rest settings are default

Corona environment slot - corona sky with intensity 1.0, model H&W
corona sun - intensity 1.0

[maru]

Ok. That explains A LOT. :)
Please revert to defaults - they are optimal and in Corona Renderer in 90% cases there is no need to ever change them. There is a "reset settings" button in the Scene tab. Is there any reason you changed them anyway?
After reverting to defaults, you can then test the different samplers.

[Tanakov]

Ok. That explains A LOT. :)
Please revert to defaults - they are optimal and in Corona Renderer in 90% cases there is no need to ever change them. There is a "reset settings" button in the Scene tab. Is there any reason you changed them anyway?
After reverting to defaults, you can then test the different samplers.

Ok reset it please, it works too good :)

[Bzuco]

I made quick test on bridge scene. Default settings gives mi less noise, which is obviously, because of low number of samples I am using in my settings, but AA was worse (13 vs. 37 total passes). I am using 192 total passes, which is enough for perfect AA with hann image filter(smaller object far away like fences or grass blades) and with very low GI samples it is working...and I am still not using denoising...so I can easily almost halved render time.

Portal fraction in int. scenes gives mi less noise at 1.0 value. In ext. it has no effect.
Embree settings...I did not notice any changes when I switch both off, but gives me more samples/s.
Adaptivity off...when it will have more algorithms in it self, I'll switch it ON :)

It is simple catching the best render time at same quality with everything what corona allowes me to change. So I agree, default settings are optimal in 90% cases...but If I try change something and it is working, then I can have some benefits.

If I have some time I can test my last ext. project with default and my settings and if default will be better I must accept defeat ;-)

[Bzuco]

Here is the result. Both render times same, 28min 24s on haswell 4C8T@3.8GHz.(default sett. 59passes, my 192passes). In both cases was adaptivity off, denoising off and image filter set to hann(for better details). Significant difference on water surface, white fences and grass in far distance.
For interior scenes will be probably better last default improved random sampler with default GI values, ...but who knows :)