Comment by cyber_kinetist
5 days ago
Editing Gaussian Splats is still a pain in the ass in the artist's perspective. Even if you can create a good-enough first try using scanned data or generative AI, you just end up with a rough draft that you cannot polish in any way. Existing mesh-based tools allow you to edit the geometry relatively easily, since they are in a higher level discrete representation rather than just a point cloud data structure.
It seems to me that there's some overheated rhetoric that reminds me of the tech-specific spin on the Appeal to Novelty fallacy [1], where people think a new tech is going to uniformly improve on an old tech, that if it isn't an improvement on every front it is somehow a "failure", and therefore if we like the new tech and we are on Team New Tech that we must defend how the new tech is an improvement on every aspect.
Gaussian splats are definitely interesting and do something things older tech is not very good at, but at the same time, it's definitely going to end up being a tool in the tool chest and not completely murderating mesh-based tech or something because they have a lot of other weaknesses, like editability. Or dynamic animation.
What I think some people may not realize is, that's not particularly uncommon. There's a really, really long line of graphical techs that do something particularly well but their weaknesses have kept them in a limited use. It's not a problem for Gaussian splats to become a tool in the toolchest; they aren't a "failure" if we're still using meshes for a lot of things in 10 years.
Mesh-type techs are the "default" for some good reasons.
[1]: https://en.wikipedia.org/wiki/Appeal_to_novelty
There’s some movement in this area to be able to surface quantize the splats but you are right, right now it’s simply just visual language and isn’t useful in the pipeline.
Extracting a surface mesh is possible, but the result is going to be really ugly (like the high-poly meshes from generative AI that are useless to artists)!
Mesh processing is a very difficult research domain in computer graphics that has been iterated for several decades, and we still don't have a good automated solution for retopology (Partly because the problem is hard to define in a mathematical way, but also since it's not a problem you can just solve with AI by throwing data and compute at it)
I used voxelization of the splats in the past so I appreciate the notes on the difficulty but this is sort of what PlayCanvas is doing here. Taking the splats, making voxels, meshing.
It's a novel approach and worked well in BIM a few years ago, though not anything real-time.
https://github.com/ziplab/VolSplat
With things like the latest dlss (extremely high quality run time .. reinterpretation), I wonder how precise mesh etc has to be now.
1. extract even a super approximate (meaning, like square edges, with some visual details) mesh from gen ai or a scan as a starting point,
2. move things around and define volumes for gameplay needs,
3. name things ("this is a Victorian house in a surprisingly good condition compared to the neighborhood it's in"), have human guided gen ai polish the things a bit more from the labels within the bounds of the gameplay required volumes,
4. let run time dlss fix the lighting etc from the rough geometry
1 reply →