We want to be able to use PointInstancer to instance objects with each instance having its own time. e.g. for creating a crowd where the instances are cards with an animated texture and we don’t want every character to move in sync. Experiments show that we can do this by making the prototypes be references, each with a different layer offset. If the animated texture is, say, ten frames long, we create ten references and assign them to the instances appropriately. This works, but I’m wondering if it’s the only way or whether there’s a better way?
That is the only way, @jerry.huxtable ! Our original, in-house PointInstancer carried a per-instance times array that allowed each instance to map its own timing offsets independently of which prototype it embodied. While that allowed us to do cool things like (if you saw it) the streaming Manta Ray scene that got demoed at Realtime Live in 2015 or 2016, what we found was that it created alot of complexity in the pipeline for being able to reason about PointInstancers, also to be able to predict how much memory they’d consume, and the Hydra implementation was also complicated and more difficult to optimize.
So we removed that feature for the new design that we open-sourced in UsdGeom. Some folks have been thinking about crowds-specific instancers (which both could address UsdSkel-specific concerns, and also leave the more common, set-dressing and FX-oriented PointInstancers free of the greater complications), though there are alternative designs being considered, IIRC.
We are doing a similar thing as well. All of our Assets are stuffed under a class prim (these are dynamically built) and we have a simple HDA in houdini to duplicate these automatically generated prims as internal references with an offset (which are instanceable). We then give the instancer in Houdini these as the prototype sources and use an inherits as the arc in the instancer.