KD tree bug - again

[CTZn]

Plain MLT fixed that way:

Code: Select all

        <metropolis>true</metropolis>
        <large_mutation_prob>0.667</large_mutation_prob>
        <max_change>0.035000</max_change>

I bet the scene could use even higher values because I could see a few pixels at the beginning, but they are fading away rather quickly. Obviously I also switched to BVH

[rgigante]

Hi guys,

by my very very small knowledge (and this should be confirmed by someone else) in a standard scene with only diffuse materials also using PT it's not admitted to have fireflies due to the lack of specular reflections (lack of accumulation points???). This means by my point of view this should be investigated as a bug.


Here my two cents.

[fused]

Hi guys,

by my very very small knowledge (and this should be confirmed by someone else) in a standard scene with only diffuse materials also using PT it's not admitted to have fireflies due to the lack of specular reflections (lack of accumulation points???). This means by my point of view this should be investigated as a bug.


Here my two cents.

thanks Riccardo

[CTZn]

Yes rgigante, by lack of contradicting sources I'd tend to adopt your point of view as well (and yours fused, by the way). I spent hours today documenting myself and nuked pixels are mentioned only when it comes to specular materials. What results is the lack of elements we users are facing for an answer.

I think you are getting me wrong, fused: I'm adopting a neutral point of view. Sadly no qualified person answered us clearly; I'm pretending to exactly: nothing

What we are qualified for, is to say that users expectations are constantly independent of the technics used for rendering, and from a technical point of view that attitude is strictly wrong. Again, there must be a distinction between unwanted, and incorrect results.

I'm not pretending to know to wich category the issue belongs to, but I dare wondering, this is it.

I hope the issue will be adressed if it can, either what a clear statement on the situation would still be a sane move imho.

Investigations: lights strenght had to be augmented by a factor 100 000 in order to get an balanced image using default camera tone-mapping, you want to check that when dealing with lighting... but that didn't fix the issue anyways .

Avoid using mn_cubic as splat filter if I may, even more if supersample is set to 1 ! Downsizing is still an admitted fix for nuked pixels (I prefer this expression, "fireflies" is suggesting a bug too much).

Also I replaced the rings geometry with a few spheres primitives to see if unwelded surfaces could allow nuking despite ray nudge distance, no chance with that (rings innocented ). See attached.

An interesting fact is that these fireflies are not killing reinhard's tonemapping, meaning they are not exactly out of range ? If that's because they are clamped already one could process them, at the cost of unbiasedness, precisely (unwanted!=incorrect). If that's the case then we are facing a developer's choice I believe (hence he is aware already).

I hope I didn't lost logic that time

[OnoSendai]

Hi Fused, thanks for the great bug report.
After a bit of investigation, it seems to be a precision issue which interacts badly with the triangle intersection algorithm used with the kd-tree. I've solved it by changing intersection algorithms.

[fused]

Hi Fused, thanks for the great bug report.
After a bit of investigation, it seems to be a precision issue which interacts badly with the triangle intersection algorithm used with the kd-tree. I've solved it by changing intersection algorithms.

Awesome. And... no problem
I'd have to thank you. Thank you.

What about the firefly issue?

[OnoSendai]

fireflies for non-bidir are still there; one thing at a time

[OnoSendai]

Have fixed the bidir=false fireflies.

[fused]

You rock. Seriously.