BDPathTracing vs ProgressivePathTrace Running out of samples?

[anchovy]

Vlado-

I thought about putting this in the problems section but I'm not sure what all is going on behind the scenes. I'm not sure its so much a problem as it is a limitation of the current implementation. I ran this setup with both the Progressive Path Tracer and the BDPathTracer. In both cases I set the total samples to the max 65k and was surprised to see that the BDPathTracer Ran out/finished in a touch over 8min, but with still a lot of variance. I stopped the Progressive Path Tracer at roughly the same point, even though it was suggesting it would continue on for several hours to hit the max samples.

Id love to know whats going on under the hood that the BDPathTracer ran out of sample so quickly, or is there another heuristic that is determining when to have the render stop. I guess this also begs the question what in the future might you do to not have the render stop until the image reaches some variance threshold?

Here is the Pathtracing version.


Here is the BDPathTracing version


Thanks!

-Michael

[vlado]

In both cases I set the total samples to the max 65k and was surprised to see that the BDPathTracer Ran out/finished in a touch over 8min, but with still a lot of variance.

It's a bug, some numbers seem to overflow. I thought we fixed it, but apparently not You could try a lower value like 30000 passes.

Best regards,
Vlado

[anchovy]

Interesting. I was also seeing a leak on another scene. Ill be happy to keep looking for issues as i play with it.

-Michael

[anchovy]

It's a bug, some numbers seem to overflow. I thought we fixed it, but apparently not You could try a lower value like 30000 passes.

Best regards,
Vlado

No luck, Ive tried an number of values but I still come up with something under 10 mins. Do you think any of these bugs may be hit on the next release? In not sure the BDPathTracer is high on the list

Thanks!

-Michael

[vlado]

If you get me a scene with the problem to vlado@chaosgroup.com, I can get you a fixed build.

Best regards,
Vlado

[Javadevil]

Thats bizarre that happens to BDPT version, I set off this scene last night and it would continue rendering all night if I left it.
I'm guessing since its an internal shot theres lots of ray bounces going crazy.
Both shots using Embree, AVX.
BDPT in 4h

PTPT in 1h8m

[anchovy]

Yep, that's what i've seen as well. Its nice someone else is playing with it a bit. I know it wont be next week, but eventually I think this the way rendering is going to go.

-Michael

[Javadevil]

Yep, that's what i've seen as well. Its nice someone else is playing with it a bit. I know it wont be next week, but eventually I think this the way rendering is going to go.

-Michael

I'm just addicted to renderers, so once I saw Vlado making a BDPT I had to see what sort of shape it was in
Once PT is down to 20 to 40mins a frame I can see it being use a lot more, it still won't suit everyones needs but for me it will.

[anchovy]

If you get me a scene with the problem to vlado@chaosgroup.com, I can get you a fixed build.

Best regards,
Vlado

Vlado-

Here is the scene file for you and anyone else to take a look at, and Id love a new build if you get to it. Its a bit large for email.

http://morphographic.com/VrayTest/20..._Metal_001.max

Thanks!

-Michael

[Javadevil]

Hey Micheal,
Your scene has 2 BPtracers in there, there both on, one is set to 100 passes, thats causing it to terminate early.

I got board after 42mins and stop it, it would have keep going.


cheers

[anchovy]

No way, ok that was a bone head mistake. Its funny ive had that happen once before but had caught it. Thanks for taking a look.

Javadevil, thats cool that you like playing with the renderers. Right now i'm just looking at the bounds of what can/cant the BDPT resolve. In a perfect world you'd like a method that can reality quickly resolve all the possible light transport paths and thus capture all the nice subtle effects that for so long we have been cheating to include. Ive done a fair number of years in vfx so I can cheat just fine, but I like to play with the toys that show you where we will eventually be, where cheating will be an aesthetic choice vs a necessity.

-Michael

[anchovy]

Subtle Differences

Heres a good example of what I was just talking about.

This is a sample image for the indigo manual using BDPT



And here is a not so remarkably similar image for the VrayBDPT. I didn't need to have it run so long, I just let it go overnight.



Now here is an image again from indigo using the BDPT with MLT.



Its dirty as hell but it does capture this nice reflection/refraction in the upper half of the sphere that is missing from both of the other images. im not sure if its a specular-diffuse-specular path but Id love to see if it can be captured. Part of me wonders if the remaining noise that we see might resolve into that element given enough time or is it just not possible to capture with BDPT. I was under the impression that with enough samples you can resolve anything

-Michael

[Javadevil]

I wonder if the bit in the top half of the sphere, if its a reflection from the caustic on the floor ?
Its amazing seeing the natural phenomenon that unbiased renders pickup.
I've been using a couple of different ones for ArchVis work. I find them to slow, but continue to use them for the quality.
It will be interesting to see the development of VrayBDPT.

Do you have that cornel box in 3dsmax ? I could easily make own but will be a bitch to match camera and spheres, etc..

[anchovy]

Do you have that cornel box in 3dsmax ? I could easily make own but will be a bitch to match camera and spheres, etc..

http://morphographic.com/VrayTest/20...re_V01_001.max

This is a slightly modified Cornell Box to match the Indigo render. If you are interested I also have an original based off the Cornell data.

-Michael

[vlado]

Part of me wonders if the remaining noise that we see might resolve into that element given enough time or is it just not possible to capture with BDPT

Yes, it will resolve, eventually. The image looks a bit odd though - if you move the camera, can you really see that caustics pattern on the box behind the sphere?

Like I said, the next major version of V-Ray will handle this case much more efficiently (not with MLT).

Best regards,
Vlado