Noise and mesh lights - Path or Branched Path Tracing?

[silex]

In my scene I have couple mesh lights, and they produce high amount of noise.
Unfortunately due to their shape they cant be replaced with normal lights.
I’m trying to find the most optimal way to render them. So far the most acceptable way is render with high sample count with traditional path tracing option:

I also experimented with Branched Path Tracing, and below result rendered with 128 AA + 1 all - took 50% less time than above PT with 1024 samples. The only thing that I’m not happy with is how mesh light are sampled - you can see it in how the brightest light is distributed - in BPT below there is a lot more noise. That said I tried to increase Mesh Light samples in BPT but it doesn’t do much aside from increasing rendering time.

Does anyone know how to deal with mesh light sampling - either in PT or BPT?
Thanks for any help.

[CarlG]

I know nothing about BPT, I always rely on analytical lights as much as possible due to noise issues. Did you try rendering out the lights in separation? Differently sized mesh lights can produce a lot of noise. Or use analytical lights for diffuse only? Are you sure the shape of the lights are that important? I cannot make heads or tails on what the image is supposed to be, let alone any specific lighting geometry. If you don’t get a better solution, sometimes the next best thing is just the way to go.

[silex]

Hi @CarlG !
Unfortunately I can’t show whole scene, and the images from first post are just small part of it.
The important difference between them is big noise difference in most brightest parts, and PT looks much more uniform and less noisy in that area. What I’m trying to do is to figure out if BPT can give me same result with shorter rendering time.
I will render better examples.

The mesh lights are visible in the scene as objects partially covered by smoke and volumes. This can produce quite cool effects like this:

The light arc is a cylinder without caps with emission shader viewed frrom the top. Unless Blender lights can be bend like that I have no much choice.

[Piter]

Did you try clamping for direct and indirect lightning?

[silex]

Will give it a try, thanks for suggestion @Piter.

[Photox]

Something is not right, at 1000 samples you should have a much cleaner image, even with sss you would mostly see the noise in the darker areas. A guy had a similarly noisy image and it was actually that he had ‘adaptive sampling’ on . Try toggling that.

[silex]

Thanks for all suggestions. I was able to reduce noise significantly. Indeed I had Adaptive Sampling on.

1 analytic light only, no clamping, 128 samples.

Reference render without optimisation. All lights on, no clamping, 128 samples.

Clamped - Direct 20, Indir 10, 128 samples.

Clamped - Direct 20, Indir 10, 1024 samples.