by default when creating a point instancer in houdini solaris, the prototypes are left underneath it, even when instanced, which creates lots of descendants which in turn slow down considerably the loading and viewing of the data for very complex shots.

Is there an easy way to go around this issue other than python scripting from scratch a custom point instancer ?

anyone tried to recreate a point instancer through python with external prototypes ?

Here's an example:
Create a sphere, a cube and a torus, plug a python LOP and set it to use this code:

This is a great example, thx for this.
For complex ecosystems I see this option as being quite useful for hierarchical scattering

maybe something to add as a feature in the current point instancer options

npetit
Here's an example:
Create a sphere, a cube and a torus, plug a python LOP and set it to use this code:

from random import random, randint
import hou
from hou import hmath
from pxr import UsdGeom, Gf

node = hou.pwd()
stage = node.editableStage()

numinstances = 20

prim = stage.DefinePrim("/pointinstancer", 'PointInstancer')
pi = UsdGeom.PointInstancer(prim)
rel = pi.GetPrototypesRel()
rel.AddTarget("/cube1")
rel.AddTarget("/sphere1")
rel.AddTarget("/torus1")

protoindices = pi.GetProtoIndicesAttr()
protoindices.Set([randint(0,2) for x in range(numinstances)])

pos = pi.GetPositionsAttr()
pos.Set([[float(x * 5), 3.0, 0.0] for x in range(numinstances)])

rot = pi.GetOrientationsAttr()
rot.Set([Gf.Quath(*list(hou.Quaternion(hmath.buildRotate(random()*180, random()*360, random()*270)))) for x in range(numinstances)])

scl = pi.GetScalesAttr()
scl.Set([[random()*2+0.2]*3 for x in range(numinstances)])

Is there a way to do this with variants? For example, to have instancer randomly choose variants of a sphere.

Sure, you just need a separate prototype per variant. The explore variants LOP is great for creating this kind of setup.