search

tommadams / tommadams.github.io

Source for tommadams.github.io
1 stars 0 forks source link

Comments for post 2010-12-23-fighting-fireflies #4

Closed tommadams closed 4 years ago

tommadams commented 4 years ago

Comments for blog post 2010-12-23-fighting-fireflies

tommadams commented 4 years ago

comment imported from WordPress Krys said on 2011-02-23 03:04:18:

Interesting, I like the way you fight against firefly. It is a very simple but efficient way... maybe ...

But don't you think that you will remove some caustics... or simply decrease the convergence speed ?

Another small question, which material do you use for the Manifold model ? which BSDF and which parameters... I will be interested to compare your rendering with mine :-) If you allow me :-P

tommadams commented 4 years ago

comment imported from WordPress Tom Madams said on 2011-02-23 11:29:18:

The technique I describe only affects pixels with high variance (i.e. those that have been poorly sampled).

Sampling those pixels more times will only improve the accuracy of the result, so true caustics will not be removed. Since the number of pixels resampled is tiny compared the the image size (around 0.01% for a 720p image), convergence speed is not really affected.

I don't know if I still have the scene description for the manifold scene; I'm kinda busy this week, but I'll see if I can dig it out over the weekend.

I got the manifold model from the forum on jorato.com (look for "manifold.rar" on that page).

tommadams commented 4 years ago

comment imported from WordPress pl01 said on 2011-02-24 00:52:28:

Thanks Tom,

Very interesting... for sure, I will give a try :-)

tommadams commented 4 years ago

comment imported from WordPress squeen said on 2016-08-27 15:47:09:

Another hack that make a big difference with fireflies as well as anti-aliasing of very bright light source near the edges of objects is to super-sample (i.e. draw a larger version of the image), and then down-sample (shrink) the image in LDR instead of HDR. The fireflies are then averaged at the clamp value.

I like your variance reduction idea. I was planning to do something almost identical myself, but had not gotten around to it. I was going to keep a running sample-variance buffer for each pixel and progressive shoot more rays at the high ones until the 1-sigma value dropped between a LDR color threshold.