(EDIT: Christophe gave me the also hint that the new topology nodes have a build in search functionality that can be used. So i spend some time to rebuild the group search on top of that. What significantly increased the performance (EDIT OF THE EDIT: meanwhile it uses the faster method from Lukas Tönne for min/max calculations without the need of sorting. sorting still used for mean though))
Without the help of the community this would not have been possible.
Really BIG CREDITS for knowledge transfer go to Cristophe Mermoud and Lukas Tönne.
A Ton (the greatest measurement unit in the blender universe) of thanks for them and You all !!
So the group statistics nodes for float and vector attributes on the point, edge, face and corner domains with the new search can be found in the link below:
Still hoping it is useful for someone until we may have have a build-in node for that. And let me know if you found some more bugs in the setups or have any suggestions for improvement.
This is cool ! I wanted that for so long !
Thanks a lot for looking into this, I don’t have something to test it right now but I’ll keep that engraved in my head until then !
Found another Bug that i introduced while having stripped the permuted index from the group. Did that to reduce the complexity of it. And because no good use case came to my mind (only for sorting, where i would use the builtin Topo Group Sort node only. Which still gives you the permuted index). So the following output should not exists it was only forgotten to be removed. And it was internally not connected at all:
During that move I accidently plugged the permuted index in the other indicies Field (“indicies min”). Which should just hold the same index for each group element pointing to the point with the minimum (for that group).
That output is meant to be used to get the the group minimas by index. See how selecting by index is done below:
Hello, first of all thank you so much for this!
I have run into issue with dense geometry, for example surface like this with 4 levels of subdivision and above, (5 on the pic) it’s skipping points created by the subdivision. After applying subd it is the same.
“Mean” is perfect, this issue is with “Min” and “Max”.
I guess it is related to the issue of low float precision when comparing mentioned earlier, so I guess it is still somehow there? Downloaded your file few days ago.
Thanks again
BTW it gets better somewhat if I increase mesh line point count to 3 in the “MO_Topo Group Sort” (with performance hit), but if I go above that its not getting better, while performance keeps suffering more.
Hmm…have been running into something similar around two weeks ago, when i was trying to build nearly the same thing to what @higgsassuccessfully found a more elegant version, while i was still searching for the bug in my setup.
When i was cranking up the density, it broke my mesh structure. Hunted it down to the Group Attribute Statistic Node.
That was also what came first to my mind. But i am not sure.
And if so, there is unfortunately not much i could do about it, because the comparison now takes place inside the code of geometry nodes itself. As the sort takes place in the C++ code of this Node:
Sorry to say, but my attempts to simplify the problem failed, because in all my tries to do so everything functioned as expected.
I’ve been trying to use the new Repeat Zone to make a more performant version of the Group Statistics Node with only partial success in that it only works for specific use-cases where I know the IDs I want to compare ahead of time and it requires a Geometry for the stats to be attached to…
Have you had a look at the Repeat Zone? Maybe there is something there you can use?
By recreating a sorting algorithm inside geometry nodes ? (Which i would not expect to be faster than the c++ sorting algorithm of the node)
Or by optimization of another specific part of the node network? (Which i think is not really worth the work because most of the time is spend in the sorting part)
The only thing i could see at the moment that the loop helps for here is to clean up the nodegroup chain in the geometric median calculation.
Next step would be to make 9 copies of that node-group and set to every domain for float and vector and then group those and have a mess of switching logic. But I’m done for today
