GaussianAnything: Interactive Point Cloud Latent Diffusion for 3D Generation | Hacker News Reader

GaussianAnything: Interactive Point Cloud Latent Diffusion for 3D Generation

nirvanalan.github.io

83 pointslnyan10 comments

It would be really neat for 3D printing and scanning. However, I don't think it'll work with functional prints given dimensional accuracy is important. photogrammetry still has a 1-up on this I believe.

Would love to get away from photogrammetry the day splatting gets good enough

There’s work done around DeepSDF that do latent space diffusion but you end with a solid geometry suitable for 3D printing.

hackernews weirdness has forced this roundabout way of replying to you about a thread from a bit ago...

i stumbled on your post about eye tracking needs for a relative dealing with MS.

i have some previous experience working with the tobii 4c and the 5 using more off-the-shelf solutions, as well as the more 'true medical grade' side of the tech, and may have some recommendations and learnings to share.

i have temporarily added a way to contact me into the about section of my profile. once i hear from you, or in about a week, i'll clear that out.

drop a line if you're interesting in discussing.

Alright, I've reached out via email. Thank you for going out of your way to connect.

Potential for indie game devs

Doubtful they'll be optimized for games. I guess if by indie you mean unoptimized then maybe...you still don't get rigs or animation.

I mean, at that point you could just make a 2d game with the input images.

If perfect topology interferes with shipping, it's overrated.

I could think of some mechanics I could add to Tentacle Typer (my indie game) if I had 3d models of the top 2000 most used nouns. I could decimate vert soup meshes into something usable for all of them in a weekend.

Rigging would take longer. Quick and dirty pass I could do it in a month. Complex animations I wouldn't even try, but Brownian motion on some bones would breath a lot of life into things for basically no effort.

Like it or not, this kind of thing will get absorbed into some workflows eventually.

The workflow is fine in theory. I just don't think the work to clean up the asset is a net gain on making it, yet.

Same, and I say this having been making 3D reconstructions of a whole street. One should not underestimate the power of current 3D modelling tools, and modularity. The latter allows you to divide and conquer a scene by creating a set of models that can be repeated, and making different materials as needed. This is the biggest gain traditional modelling has. Not only does this make the art asset creation faster, it makes a very optimized scene.

Reply on news.ycombinator.com