very cool! it looks promising (though i am not an expert here either 😅). compared to the proxy mesh + solari re-lighting method, this would generalize better for 4dgs/dynamic 3dgs (where gaussians are not bound to a morphable proxy mesh)

i'm curious how shadows would be cast/received.

i wonder how reasonable the sdf from 3d gaussians is (instead of co-train: https://arxiv.org/abs/2404.09412). normal rendering intermediates would be good to inspect.

it may be time to retire the icecream ply 😃 I'd imagine trained scenes or better ply objects will have higher quality sdf/normal estimation.

it's an awesome direction!

@mosure slightly off topic, but since you mentioned example scenes, could the crate perhaps make use of the kew networked assets in 0.17? Would prevent having to bundle large files in the repo