CoolColJ's test renders

[CoolColJ]

I did some tests with the older build of ERPT - Apparently ERPT has been updated recently

ERPT has always puzzled me, but today I got a good feel for it.

Tried these -

Bidrectional
32 depth all round
Random Sampler
Hilbert pixel sampler
2 samples per pixel

I tried chain length 8 and 16 so far for 15 mins each - the lower number means caustics don't develop as quickly, but it seems to render bigger chunks of the image with Hilbert on a semi-tiling basis when using a 3 sec screen update.
So it covers the whole render and develops quite quickly. Reaches higher samples per pixel much quicker.
First pass is sorta darker, and gets brighter every pass in a bucket rendering fashion.

If you use Vegas or Low discrepency pixel samplers it looks like a mess!
Tile and linear is not bad, it goes down in a scanline fashion, but is slower than Hilbert

All rendered in 15mins - different chain lengths. Lower chain length = higher samples per pixel, more refined, but lower caustic developement, also not as bright...

[CoolColJ]

MLT version rendered for a long time - had to crank max rejections up to around 10000 to get all those refracted caustics and highlights on the spheres. The default of 512 not really enough - but it's not always predictable this MLT....
Can't seem to get these with SPPM. The ones on the monkey ear it can do.

Bloom added - I honestly don't think all of the possible caustics and refracted reflections have shown up yet....

[Meelis]

Interesting, looks like with low chain length it gets stusk to same paths.

[CoolColJ]

Interesting, looks like with low chain length it gets stusk to same paths.

Funny that the red sphere caustics rendered really early and smooth though, in all cases...

[jeanphi]

Hi,

It is expected that caustics will show up early through the red sphere, because it's large and directly in sun light so the bidirectional path tracer can easily resolve them.
With ERPT, the number of samples per pixel is slightly misleading because it only counts seed samples and not mutated samples, so the higher the chainlength, the more mutations and the less seed samples in a given amount of time. In case you're not aware of how ERPT works, here is a resume: it traces a seed sample, then compares the intensity of the result to the mean picture intensity to determine the number of chains, then for each chain it starts with the seed path and mutates it chainlength times, accumulating an equal fraction of the initial intensity each time.
That also explains why a lower chain length gives a better coverage of the pcture but has a harder time resolving caustics.

Jeanphi

[Meelis]

Hi

In theory is it ok to keep the pixelsamples 1 for the erpt?
BTW HUGE speedup compared with 64 pixelsamples

Code: Select all

Sampler "erpt"
   "integer pixelsamples" [1]
   "string pixelsampler" ["vegas"]
   "integer chainlength" [512]
   "string basesampler" ["random"]

[CoolColJ]

Ok that might explain why I see more caustics in all areas of the render with chain lengths > 100 even though Hilbert renders in a tile based fashion.
It's like the center of this particular render and the objects themselves do not get rendered for a long time while the caustics and indirect lighting develop long before hand.


meelis - Yes low pixel samples is much quicker, but I would not use Vegas, just too random and messy looking, try Hilbert, the image looks more coherent in the early stages.

[jeanphi]

Hi,

Since ERPT mutates the samples, it is not that important to have low discrepancy so you can use the random sampler with a low number of samples per pixel. Using the Hilbert pixel sampler, you'll have a moreaccurate render on the sampled areas but a bad overall look, while using the Vegas pixel sampler you'll get a fast overall look, but you'll need to wait till the end to refine the picture.

Jeanphi

[CoolColJ]

ERPT render with chain length of 100 - rendered for 4 hours, still not that clean and not all caustics are shown. Will try with chain length of 1000 next
Definitely slower than MLT - the above MLT render was less than 4 hours

Those funny ghosted pixel hotspots you see when you start the render don't completely go away.
Note the funny artifacts on the edges - I think it maybe due to my filter settings? I used MN with blur of 0 and ring of 0.5, 2 radius

edit - I think it's a ERPT glitch, because it still does it with different filters. Maybe Hilbert pixel sampler glitch?
Tried linear, low discrempcy and tile pixel sampler and all have the same glitch..... It's like a halo around all the edges of the objects

[CoolColJ]

Direct light sampling in SPPM is not very accurate....

SPPM settings used - store on Glossy (blue cube), radius 0.1, arc 1.00 (no point in shrinking if it's already small), 1 millions photons, AMC (needed), Random and Vegas pixel sampling all round
MLT render above as reference

No direct light sampling - looks like the MLT render

With direct light sampling - the lighting looks strange and the caustic of the red sphere is completely different and smaller!