Brand new way to kill hot pixels

[kivig]

Thread
Some sample:

Before:

After:

[Chris_TC]

The litle imprecision of both these technics ( masking and darken blending ) is the fact that in very noisy parts of renders ist the result a bit darker as the corect result.

Darken really acts as a mask, it doesn't produce a darker result than would be correct.

Imagine two pixels on layer 1: RGB 60, 60, 60 and RGB 255, 255, 255 (the hot pixel)
Imagine two pixels on layer 2: RGB 255, 255, 255 (the hot pixel is on the left in this render) and RGB 75, 75, 75

Layer 2 sits above layer 1. If you use darken on layer 2 it will not replace the left pixel (because 255, 255, 255 is brighter than 60, 60, 60). However, it will replace the right pixel (because 75, 75, 75 is darker than 255, 255, 255).

You will end up with the correct two pixels: RGB 60, 60, 60 and RGB 75, 75, 75. The hot pixel of either render does not show up in the result at all.

[kivig]

Darken really acts as a mask, it doesn't produce a darker result than would be correct.

Actually it does. Every of two pictures has random shift from what's "right" (noise) up and down. And darken takes only the lowest values. Easy to check - take two very noisy fast rendered images and compare to single enough sampled.

Btw anybody knows where do hot pixels come from?
I suspect they are not a defect at all. Just too low sampling for some small, bright object. If I leave render long enough I always can find the source of that light and it seems logical.

[sam75]

so you are all using rayepsillon to make variations ?

[Chris_TC]

Actually it does. Every of two pictures has random shift from what's "right" (noise) up and down. And darken takes only the lowest values. Easy to check - take two very noisy fast rendered images and compare to single enough sampled.

I stand corrected, no need to try it out because what you say makes perfect sense.

Can the program you posted be used for animations?

Btw anybody knows where do hot pixels come from?

This I'm curious about as well.

Strangely enough, the longer you render the more fireflies seem to pop up. I had rendered an image for 1 hour and saved it to disk. Then I let it render for an additional 30 minutes and saved that to disk. Upon comparing the two images, the one that rendered longer had considerably more fireflies.

[tomas_p]

Exist a way how to remove darker parts from the result. You need not 2 renders but 4.

You need 4 renders wit other noise and each with 25 percents of desired number of samples (A, B, C, D). Make two couples (A, B) and (C, D). At each couple apply the masking proces and at these two results (AB_result and CD_result, now both without fireflies) apply the same proces, only the mask must be invert (you need remowe the darker pixels not the brighter).

The one problem of these methodes is how to make 2 or 4 renders with other noise without rayepsilion changing. As the rayepsilion changing can make problems.

[kivig]

Can the program you posted be used for animations?

Yes. It's not tested well yet though. There's some sort of instruction if you follow the link.

Edit: It has to be frame sequence (png, jpg, tif, bmp) with six digit frame before extension (like Octane creates automatically).

[sam75]

Exist a way how to remove darker parts from the result. You need not 2 renders but 4.

You need 4 renders wit other noise and each with 25 percents of desired number of samples (A, B, C, D). Make two couples (A, B) and (C, D). At each couple apply the masking proces and at these two results (AB_result and CD_result, now both without fireflies) apply the same proces, only the mask must be invert (you need remowe the darker pixels not the brighter).

The one problem of these methodes is how to make 2 or 4 renders with other noise without rayepsilion changing. As the rayepsilion changing can make problems.

what kind of problems ?

[tomas_p]

hi sam75

what kind of problems ?

Problems like this...

On picture below is a bottle with shaped label. The label si maped on tubus with bitmap controled transparency. If is rayepsilion 0,0000 it works good, if is 0,0001 the material unedr tubus has no samples on large area. Or if you have doublesided glass, you use hdri background and the rayepsilion is to higher, the view behind the glass is is wrong (scaled or shifted)

[sam75]

hi sam75

what kind of problems ?

Problems like this...

On picture below is a bottle with shaped label. The label si maped on tubus with bitmap controled transparency. If is rayepsilion 0,0000 it works good, if is 0,0001 the material unedr tubus has no samples on large area. Or if you have doublesided glass, you use hdri background and the rayepsilion is to higher, the view behind the glass is is wrong (scaled or shifted)

thanks for the explanation, could you try by changing rrprob value (eg 0.1) rather than rayepsilon ? It works too but I don't know which one is better.