Hacker News Reader

This field is moving fast! Something like this has just been released. Checkout DreamFusion, which does something similar: They start with a random 3D NeRF field and use the same diffusion techniques to try to make it match the output of 2D image diffusion when viewed from random angles! Turns out it works shockingly well, and implies fully 3D representations are encoded in traditional 2D image generators

https://dreamfusion3d.github.io/

There's a bunch of NERF tools that can get pretty close to good 3D assets from static images already.

Yeah, I've been starting to explore those. Its all crashing together quickly.

you can already do this, just not in real time yet. You can upload frame sequences to Polycam's website for example, but there are several services out there which do the same thing

With this you can do it with things that don't exist. I'm excited to explore the creative power of Stable Diffusion as a 3D asset generator.