Looks cool. I teach 3D modelling. I would love to see the meshes this approach produces. I'm gonna guess that they are poorly topologized. Agreeable topology enables ease of texturing, editing, animation and dozens of other things. If this could be generated then I may well be out of a job in a few years.
For any heightmap based 3d models this can easily be done with customised image diffusers. It’s hilariously trivial to implement for terrains, human faces or anything that relies on grayscale alpha maps.
It's just been quite useful in my own R&D work, so thought I would try it a few times today to connect the discussion to actual primary sources from research.
But yes the message was partially AI generated. Still, thought it would be constructive.
I played with Shap-E when released. It did download a model on first run, but the actual processing happened locally.
Does this embed the model in the repo? Or does it still need to download the model from somewhere at some point?