Reasons for Fireflies/Hotpixels

[Vue2Octane]

As I understand it, fireflies will be always be there in a path tracing algorithm and are the extremes of the standtard deviation o the monte carlo solution. But anyway, I find removing them is not so difficult. For me Octane does a pretty good jon, by first picking the extreme whites out of a 'non-white' area and then just replacing the single pixels by next neighbor. I think this is what happens when you apply hot pixel removal down to 0.7 or similar. After that they seems to apply blurring.
But the fact that fireflies are really only visible on backgrounds that are themselves not white, and that they show as really sharp peaks, make identifying them easy. Once you got them, replacing with next neighbor works quite well, because it is only one single pixel.

[Vue2Octane]

As I understand it, fireflies will always be there in a path tracing algorithm and are the extremes of the standard deviation of the monte carlo solution. But anyway, I find removing them is not so difficult. For me Octane does a pretty good job, by first picking the extreme whites out of a 'non-white' area and then just replacing the single pixels by next neighbor. I think this is what happens when you apply hot pixel removal down to 0.7 or similar. At stronger hotpixel removal they seems to apply blurring.
But the fact that fireflies are really only visible on backgrounds that are themselves not white, and that they show as really sharp peaks over only one pixel, makes identifying them easy. Once you got them, replacing with next neighbor works quite well, because it is only one single pixel.

[glimpse]

It's probably already a lot been said (here & in the other threads)..
FF is a side effect of unbiased engines..& we have to get used to that
or find a way to deal with =) Just for the record, as it's on the same topic,..

I've tried to come with a simple scene (without faking the light using like DL kernel),
I've wrote a bit about what was tweaked, done & why: gains & downsides..
If You interested You can get it here: Lounge Chair - Free scene

[indexofrefraction]

you cant eliminate a small light source if you need a candle.
you cant turn of glossy surfaces if they need to be glossy
so the conclusion is: you cant avoid the flies?

as from what i read until now:
- use sufficient image size and very low hotpixel removal value, maybe scale down later
- use pmc (?) needs to be tested, caustic blur to 1 is needed?
- make 2 renders with some values changed (glossy depth for example) then join them in psd by "darker color"
- use neatvideo or other noise reduction plugins

hm and if you follow dsyee or grimms explanations…
shouldnt it help to raise the ray depths? (diffuse/spec/glossy/etc) to get more accurate samples?

[dsyee]

you cant eliminate a small light source if you need a candle.
you cant turn of glossy surfaces if they need to be glossy
so the conclusion is: you cant avoid the flies?

Not necessarily...for example there are "tricks" you can use to emulate small lights. In this scene the only illumination is from the "bulb," which I surrounded with a box with opacity = 0.

If I set the bulb to be the emitter, it looks like this:

PT 2000 samples, bulb emitter

If I set the bulb to be white diffuse, no emission, and change the box to be the emitter, it looks like this:

PT 2000 samples, box emitter

There are differences, but depending on what you need to accomplish, you can perhaps increase the polygons in the box to more closely resemble the bulb. But in general, emitters with flat surfaces are less noisy, so fewer polygons = better.

PMC with the bulb as emitter also results in fewer fireflies, but it takes longer to reach 2000 samples.

Here's a wireframe of the emitter setup so you can see what's going on. I rotated the box in order to prevent the emitter surfaces from being parallel to the walls:

Emitter wireframe

[indexofrefraction]

hm interesting reply….
in my scene i have 2 lights using a light-object with octane-light tag.
the geometry is set to sphere (5x5x5cm) in those lights and maybe the generated mesh is not optimal.

so i did the following test version:
instead of the light-objects described above i used 5x5x5cm sphere meshes with an emitting material.
i used shere-objects, with just 6 segments, set to icosahedron (so that all faces have the same size)

sadly close but no banana, this second setup showed just a small difference in noise/fireflies :/

[dsyee]

Too bad - maybe if you posted an .ocs or a render, folks could give you more specific advice?

[indexofrefraction]

Hurray, I solved it!!!
below is my picture having tons of fireflies, i post it for others having similar problems!

the problem:
the five spotlights in the ceiling were 5cm above the ceiling (thickness maybe 10cm, with modelled holes)
result was a mess of fireflies getting more and more with rendertime / increasing samples!

the solution:
pushed down the spotlights, so that they are exactly in the surface of the ceiling -> the flies were gone!

phew, i'm glad that it was a setup error, i was already doubting octane…
so if you encounter problems like that, check all your lights in detail, one after the other.
if needed disable one by one to find the light causing the issues!

[indexofrefraction]

and to visualize that :

[FrankPooleFloating]

Funny IOR, I think that was the very first thing I posted about, way back when I started Octane over a year and half ago. My first test render was a room with recessed lights. As soon as I put emitters at same plane as ceiling, the FFs went bye bye.

I think I asked this before, but can't remember if it was ever answered: Are FFs sometimes floating in empty space and not necessarily on surfaces?... Sometimes it really seems like some of them are.. And IF they are, you would think that they could somehow be killed programmatically... test for if pixel is on a surface or something... I dunno....