interior test [Distributed Path integrator explained]

[NiZu]

Hi, so , we could make a top 3 list of things that create troubles with LDhilbert + DP

3rd : scene geometry complexity .
Not just polycount , e.g. even a 50k quads scene with things like niches , long curve corridors , porches , will make complicate lighting paths.
2nd : materials
matte translucent , perfect matte , strong bump (i think)
1st : lighting : sun or many small, very bright artificial lights.

Lighting seems the real problem. the other 2 cause slowdowns which are ..absolutely normal

this is an interesting comparison : the bright yellow spot on top is sunlight ,
the remaining light is rectangular mesh emitters.

so 'dumb' samplers solve easy light fast ,
metropolis is slower but can solve difficult ligh sources.
What makes sun a difficult lamp is the 'penumbra' area : where a dumb (random) sampler sees only few rays but this rays have still an important contribution.
Metropolis takes that into account , but on the other hand takes many samples to converge (even for simple lighting)

Might be useful to distinguish that noise (in the hilbert test) from 'fireflies' caused by reflective caustic , which are smaller ,less frequent and generally brighter
(one way to distinguish them is using the maxwhite tonemapper : with fireflies the rest of the pic will be pitch black , if it's just noise you'll still see some of the remaining light)

[NiZu]

Edna : Nice lamp !
It's a different situation from the interiors i'm testing. i haven't really tested translucent yet, but indeed you might need to have more reflection diffuse samples than glossy . since that scene has a predominant diffuse material (an expensive one , since it's translucent)
I don't think Metropolis makes DP tweaks less important .. MTL finds small lightsources , DP is used to focus on the most useful ray 'types' ..i don't see a conflict
(though i don't know mtl well)
I guess , MTL would be crucial if your emitter sphere (inside the box) was very small : a dumb sampler (like LD) would 'miss' it most of the times. (=noise)

Atom:
i made a quick test with History museum : that gave me the idea of placing 'geometric complexity' as least problematic in the top 3 list (previous post) :
this is a 4 minutes render (800x600) with the easiest possible lamps (big , bright quads ) LD+DP (setting similar to yours)

It doesn't mean much that this setup rendered fast ..but shows it's not polycount / geometry itself .. rather it's that complex scene take long as soon as you put any complex light source in them.
(Anyway , the spots were probably the most important factor in your setup (MLT vs hilbert thing, previous post)

[edna]

In a similar vein of exploration, that balance of samples is very important. You have said it already, but I thought an example would help. In the image below you can see direct light samples of 1, 10, 100. All stopped at 2 s/px, taking 4, 4 and 7 mins respectively.

Massive difference in noise between 1 and 10 samples, but no change in time. Little difference in noise from 10 to 100, but almost double the time. It's a pretty small test, but it makes the point quite well I think.

By the way, it isn't a lamp but a 4m high bat house in London. Construction of the real thing is almost complete, so I am just having some fun with the model.

[Atom]

I thought I'd give higher samples for the Direct field and boy what a difference.

I think I found the "realism" button. By setting the hilbert samples to half the direct samples it really cleared up everything, fireflies and all. It certainly is not fast, but it is clean. This is important to me, because now I can at least know how to setup Lux for a final.

After 10.32 hours, 532% efficiency @ 85.42 samples.

The only problem I have with this image is the glass. I purposely set refraction to lower than the recommended. I thought I could get away with 2 bounces, but I think I may have needed at least 3.

[jeanphi]

Hi,

Do you use normal glass or architectural glass? I think you should enable architectural if that's not already the case. As far as I remember DP isn't fully optimized yet for architectural glass, but it should work better.

Jeanphi

[NiZu]

Edna :
the bat house project is beautiful !(i searched google)
Test is very useful , but a bit scary that 100 samples aren't much better, sounds difficult to find a generic reccomended setting .
Atom:
I will try that , in effect the noise in your previous post was in direct light ..might help with sunlight also.
With 10 hours it begins to be a problem to compare (with MTL+DB) might be useful to launch 2 renders togheter (2 cores each) using just a small crop of the pic (blender border)
Jeanphi:
i've always used arch glass , i haven't noticed anything strange..(but i can't tell about rendertimes)
edit: i mean i don't know if arch.glass is faster with other integrators , it sure is faster than glass with DP , and never gave me any problem.

[NiZu]

There's something i've used in the first test scene but i haven't mentioned explicitly , because i feared it might be wrong and introduce errors , and i found out it does
(i used it also in the comparison mlt vs ld : but it doesn't changes the resullts , i've checked )

I saw that invisible lights aren't possible , but it's possible to have light go thru an emitter: using an arch. glass material with white refractivity , black reflectivity , emission on.
The purpose was to have light panels on windows (fast , but boring result) and sun (slow but realistic) combined.

What happens is: (i've made some small tests) The light goes through arch glass unaffected (it doesn't coun't as a refraction) , so no problems/ errors with direct sunlight.
But here's the problem : then light bounces off the scene, hits glass emitter again , then back to the scene ... and i got some fireflies , shadows (i guess because it saw the glass emitter which isn't bright as the sky and doesn't reflect)
So, right now this doesn't seem a good method ..definitely not suggested , i guess

[Atom]

I am using architectural glass. I have a high IOR, however. (1.41)

I did another test with more samples allocated to refraction but it did not change anything. I also reduced the hilbert and direct samples a bit.

Here is the image after 5.5 hours. The last one at 10 hours was simply when I terminated it, but it does look like the image becomes usable much quicker than the 10 hours.

5.5 hours, 53.29 samples with 452% light efficiency.

The 10 hour and the 5 hour renders look pretty much the same to me.

[jeanphi]

Hi,

I think that you don't have enough bounces for glossy. The number of bounces isn't the number of bounces of this type in the path, but the the maximum path depth at which this kind of bounce will be configured.
So if a surface is behind a glass plate , you already have 2 bounces and you are at bounce 3 at your surface, so if you want the path to leave you need to configure at least 3 bounces.

Jeanphi

[NiZu]

Hi thanks for the explanation about glossy depth, i'm not sure what problem it produces though , the noise on glass ? (never had similar problem) or darker glass ?
(happened in my 1st scene but was barely noticeable , for me, an acceptable sacrifice for speed.

Also 5 hours is a good time , Atom what have you used for sky ? are your windows emitters ?
i'm doing some tests with history museum .. using mesh emitters (blackbody unfiform texture) : i've replaced the glass panels (that had thickness !) with single sided planes ..and the difference is huge : 32 samples in 1 hour (clean pic in 2)
I knew that the thick panels , with a back face shooting light outside , were not efficient .. and indeed now it's 3x-4x faster !
So, at the end, this scene is 'in line' with the rest... a little heavier as you'd expect, 1,5-2 times slower than haunted hallway and the Buddha's room .
I'll render with exr map on windows (=light material) let's see how long it takes...