We run into this quite often! This is a function of the training data. Renders usually look smooth due to interpolation of normals, but most model “learns” this faceting when converting the mesh to SDF. A pre-processing subdivision pass or PN-triangles might help here

This has been my experience as well. The number of “one off” tools I’ve written in the past few months have increased my productivity significantly; even with the current LLM limitations. I’m surprised many experienced devs still remain “AI skeptics”

This is the reason I stick with gltf (although it has its own set of problems, like you’ve pointed out before). I find it’s much easier to understand and extend. I’m surprised an in-between good storage format doesn’t exist! Maybe we need to make one ;)

The library dependencies from usd is the major reason I avoid it. There’s github.com/lighttranspo... but I haven’t given that a full evaluation

What’s your format of choice for 3D? USD? glTF has the big advantage that it’s easy to understand and extend and has broad visualization tooling almost everywhere; but I agree, it’s a shame a transmission format is used for storage

Beautiful work Nick (as always)! I’m curious on your thoughts if this can hurt generalization, where the model represents a simple/smoother latent manifold but biased towards low-freq dominant modes and unable to capture high frequency details? But maybe 3D has enough structure for it work well?