Magic3D: High-Resolution Text-to-3D Content Creation | Hacker News Reader

Magic3D: High-Resolution Text-to-3D Content Creation

research.nvidia.com

94 pointsT-A15 comments

The current results are... whelming, but this technology could be an invaluable boon to indie game developers and hobbyists. It may not be long till it's viable -- at least for rapid prototyping stages.

I think large studios would benefit from this too. I know some that use a workflow pipeline where artists are tasked with creating themed assets to build a catalog, which another team will use to build campaigns and quests for a specific MMO (forget the name).

Perhaps at the same cost, they would be able to build such a catalog quicker, especially if they need less hyper specific things in bulk.

Something tells me that these large studios might not be so large anymore. Hopefully this means more smaller AAA studios

Similar work with impressive results: https://bluestyle97.github.io/dream3d/

This is far more important than it initially appears. Current AI image generators let you create pictures of your wildest dreams and desires. The logical conclusion of this is the ability to create perfect VR worlds of your fantasies and enter and explore them. This is truly the beginning of The Matrix.

Truly untapped, the future is

https://www.youtube.com/watch?v=g1HHaJ-ILXo

Maybe it’s nothing but I expected to see some wire frames of the meshes as a way to see the mesh quality.

If I understand it correctly, it is reconstructed from multiple diffusion-generated views via something equivalent to photogrammetry, and then further refined, so I doubt the topology is much different from something you'd get from Reality Capture out of the box.

It likely doesn't generate good edge flow for symmetrical meshes, with hand-picked loops and poles. It's possible we get something like that eventually, but I suspect it's at least a couple of years away.

So probably good enough for static models (perhaps with decimate or other modifiers as needed)? But not great for something you'd need to rig and animate, or something that requires tweaking after the model outputs it?

When you click the buttons under an example often enough, you can download the mesh file for it. It's not the best UX, but the data is there.

If people has missed it, NVIDIA also recently released this paper that could be used to create 3D models: https://nvlabs.github.io/genvs/

This is really neat but very low-poly. It needs more granular control. "Make every other archway larger. The rotunda should have a sharp edge. Make this section twice as long as the overall width."

That background music tho…

Not to mention the video style itself, which appears to have been portalled in straight from the 90s.

In most examples the Imagen examples are actually closer to what I'd expect to see as well. Less detailed, yes, but a better and more pleasing result overall.

I expect these kinds of single chunk meshes are not that useful. Things will really get interesting when someone creates a model that can generate 3d objects from individual primitives.

That's cool and all. But what I really want:

https://www.youtube.com/watch?v=R2IJdfxWtPM

Reply on news.ycombinator.com