5
$\begingroup$

I read through the conversation on this question: Is there a way to render a single frame using network rendering in cycles. I tried the image stacking method that was suggested and got pretty good results. I rendered my image five times, each with 150 samples and a different sampling seed. I then rendered the same image with 750 samples to compare. (In case you're curious, the 150*5 stack had slightly more noise than the straight 750 image, so you are sacrificing some quality by splitting up the work in this way)

Then I wondered how much of a quality increase I could get if I added the 750 image to the stack, which brings me to my question. How can I reduce noise by image stacking if the images don't have the same number of samples?

So my first five layers have opacity 100%, 50%, 33%, 25%, 20%, respectively. These are the layers with 150 samples. If I wanted to add my 750 sample image to the top of the stack, what would be its opacity?

$\endgroup$
1
$\begingroup$

Short answer:

My guess is that it should be 50%. The combined samples of the images above it is ~750, which is the same as the number of samples in the rendered image. Since the rendered image and the combined images are more or less equivalent, They should be combined with a 50/50 mix.

Long answer:

Each image's opacity is based on how many samples it has compared to the image above it.

E.g. if you render 5 images with 10 samples each, the second image should be mixed with the first with a factor of .5 (or an alpha of 50%). This will effectively result in a 20 sample image (as you pointed out, it's not quite the same as 20 rendered samples, but it's reasonably close), so the second image should be mixed with a factor of .33 (alpha of 33%) and so on.

Since the resulting combined image is basically a 750 sample render, it's the same number of samples as your 750 sample render. You can just tack it on top with a transparency of 50%.

$\endgroup$
  • $\begingroup$ Yes, sure, but I think OP's point is to consider that five 150-sample images are of lower quality than one 750-sample image - that is, what is the sample equivalence of five 150-sample images? Once you know that, the math is easy, as you've outlined. $\endgroup$ – wchargin Nov 7 '13 at 5:58
  • $\begingroup$ @WChargin Any ideas on how to measure noise? If we could do that, finding the sample equivalence would be easy. $\endgroup$ – Justin Nov 12 '13 at 20:13
  • 1
    $\begingroup$ @Justin See stackoverflow.com/q/8960462/2730823 $\endgroup$ – gandalf3 Nov 12 '13 at 20:17
  • $\begingroup$ I've been doing some experiments, but I'm not sure that there is a difference in noise.. it's pretty subtle (I've just been eyeballing it though, so not very scientific) $\endgroup$ – gandalf3 Nov 14 '13 at 3:10
  • $\begingroup$ I came across this forum thread on measuring noise in Cycles. It wouldn't quite work in this situation since it compares multiple samples from the same render, but it's pretty cool and might end up in Cycles some day. $\endgroup$ – Justin Jun 6 '14 at 19:41

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.