SOMEONE JUST FOUND A WEIRD WAY TO GENERATE 3D MODELS WITH AI 🤯

And the clever part? The image generator doesn't even need to understand 3D geometry

Here's the experiment:

Angel trained a Krea 2 LoRA on voxel renders from just TWO viewing angles

Then trained a tiny model from scratch to reconstruct actual 3D voxel models from those images

No giant text-to-3D pipeline involved

The results are still rough, but imagine this approach with much stronger image models

Generate two consistent views → reconstruct a 3D asset

Could be a pretty interesting shortcut for game assets and procedural worlds

This is the kind of weird AI experiment I love seeing
https://x.com/Angaisb_/status/2121544077418139723/video/1