16 comments

[ 3.7 ms ] story [ 62.8 ms ] thread
The mustache beefed up to fill in for Homer’s nose in profile. Mwah.
This is very cool. Once we can generate good 3d models (possibly even rigged ones) we're so much closer to ai-generated videogames.
Not so fast. These aren't game ready by a long shot. The topologies are crude and inefficient, and moreover they do not deform with classical animation tools.

There's clearly a lot of promise, but we're still a long way out from this working.

I totally agree, hence the "once we can.."
Don’t dismiss it just yet, all it takes is a retopo with a 3D artist and you have something game ready.
And there are some great retopo algorithms coming out in the last few years that could automate this, my favorite being one that uses physics simulator to disperse a net across the model surface. I would imagine there are even AI driven ones.
Isn't retopo something that an ai could do well?
Yes but an artist should fix seams and small topographic errors that may occur.
The closest thing I can think of is reallusion. Used it for a project once a few years ago for generating a model from images and audio to lip animation. It was okay, but not great for real-time audio as it kept crashing. It might've gotten better tho since
What’s on the back of Dali’s head?
"A DSLR portrait of..." means nothing and if the phrase influences the output of a generative model, in a way different from "A photograph of...", that implies the model has learned something stupid.
> "A DSLR portrait of..." means nothing

I'd agree that DSLR vs. mirrorless isn't a significant distinction but DSLR vs. smartphone suggests possible differences in format, technical quality, bokeh, etc.

I was confused by this too. The photography term seemed so out of place to me that I assumed this was some new other meaning of DSLR, like deep spherical latent representation or something. but Ctrl+F didn't find anything.