Adaptive sampling improvements

[lacilaci]

One more thing.

Can this whole adaptivity be tweaked to prioritize undersampled areas even before noise?

look at this image:

to me this looks like noise is picked up ok, but there are areas that are not noisy, but more like almost not even sampled/completely black. So I don't know if this makes sense but if this could somehow force luxcore to first push at least some samples where are the super dark areas it maybe could be better...?

Now, eventually some samples end there and adaptivity picks it up

But I'm thinking if it could speed up rendering some more if those parts would be somehow first to render... Not sure if this makes sense, just a thought. Overal I really like this improved adaptivity now!

[marcatore]

is this feature already in the daily BlendLuxCore build?If yes, it's just using the adaptive settings? right?

[B.Y.O.B.]

is this feature already in the daily BlendLuxCore build?

In about 20 minutes when the build is done, yes.

If yes, it's just using the adaptive settings? right?

It is active without any user interaction. When using adaptivity > 0, the new behaviour will be used.

[alpistinho]

to me this looks like noise is picked up ok, but there are areas that are not noisy, but more like almost not even sampled/completely black.

I don't what setting you were using for step size and warmup but what is happening is that the metric is considering that after (warmup+step) samples, this region is closer to convergence than the rest of the image. Up until (warmup+step) it should have received just the same amount of samples.

After that every step samples the current convergence will be reevaluated according to the result of these last samples.

That means that this region being undersampled is simply the result of the metric we're using. I couldn't find anything reasonably simple that looked better though.

If we somehow changed the sampler to always shoot samples if the region is super dark like this, we would probably harm other scenarios.

Now, eventually some samples end there and adaptivity picks it up

As you've noticed, after the other regions get closer to their own convergence, this region starts getting samples.
The adaptiveness is not the same in the time domain. It tries to keep the noise constant accross the image, instead of keeping the amount of samples constant.

It is interesting putting both warmup and step as 1 and seeing the adaptive jump around the image, much like metropolis does. In the tests I've done it ends up converging to the same result as using a longer step size

Overal I really like this improved adaptivity now!

Glad to hear

[zeealpal]

Doesn't setting the adaptive strength to a lower value achieve the same pretty much; If its 0.7 then are 30% just sampled across the image and 70% use the adaptive map?

Feel free to correct my understanding it's more of a guess

[alpistinho]

Doesn't setting the adaptive strength to a lower value achieve the same pretty much; If its 0.7 then are 30% just sampled across the image and 70% use the adaptive map?

Yes it would more or less.

As future reference, here it is the complete flow with some small steps ommited:

• During (warmup + step) samples the sampling is uniform

• At this moment, the current image is saved as reference

• After step samples the current result is saved is compared to the current reference

• The convergence metric [1] is calculated per pixel, as in a map

• The values are filtered by a 2D box filter to make the adaptiveness smoother in the space domain

• The mean and standard deviation for the map is calculated

• The standard score [2] is calculated for each pixel metric, clamping it to 6 standard deviations

• The map is normalized according to its max and minimum value so all pixels are in the 0-1 range (that will be the likelihood of being sampled every time this pixel is considered. Also is the content for the NOISE channel)

• The current image is copied as reference

During sampling, the Sobol/Random samplers will go over every pixel and draw a random value in the 0-1 range. If the drawn value is smaller than the NOISE value for that pixel, it is sampled. If it is higher, it is not sampled.
The adaptive strength is used as the minimum likelihood a given pixel can have of being sampled in a given sampling attempt. (A strength of 0.95 means that every pixel has at least 5% chance of being sampled.)


[1] Dammertz, H. A Hierarchical Automatic Stopping Condition for Monte Carlo Global Illumination
[2] https://en.wikipedia.org/wiki/Standard_score