Value of photo realism for simulated training images

I’ve successfully used nVidia Isaac Sim to create useful training images for object pose estimation. Fine.

Now I’m evaluating whether the added photo-realism from Isaac Sim adds demonstrable value/accuracy to my ML results - over using more less photo-realistic (clunky) output from simpler simulators (with less intensive hardware requirements, e.g., say webots)

Yes, in all cases I’ll do color dithering as part of augmentation (and whatever else makes sense, e.g., flips).

Any thoughts? (I haven’t found much in the literature that touches on this.)

Thx!!! p