7 ms·
The leading 3d gen techniques are latent diffusion models that work off voxelized intermediate representations of objects. When you convert these into meshes, v
by LarsDu88 1mo ago
The leading 3d gen techniques are latent diffusion models that work off voxelized intermediate representations of objects. When you convert these into meshes, very often they will have poor topology. Topology is important if you need to modify these meshes or animate them. Bad topology and excess verts is also bad if you're building assets for a game.
There's still more refinement of existing techniques that need to be done before AI generated 3d models match what humans produce.
- site-packages1 1mo agoMy experience here has been so bad with the topology that I usually use the created mesh as an example and rebuild the whole thing from scratch to reduce the verts. Not even worth the time.
- LarsDu88 1mo agoThe real algorithm that's needed is an intelligent AI mesh retopologizer. Honestly more of a reinforcement learning AI problem than a generative AI problem IMO
- pennomi 1mo agoA good auto retopologizer (AI or not) would be one of the greatest possible innovations in 3D. Add to that better unwrapping and rigging, and you’ve eliminated most of the mechanical grunt work in 3D, leaving the truly important sculpting and animating to the artists.
- LarsDu88 1mo agoI'm sorry to say this, but the voxelization diffusion model I mentioned is effectively AI sculpting, and AI animation is already quite good (for humanoids at least) Virtually all the dominos have fallen except good topology at the moment, and unfortunately that's the biggest domino there is. AI videogen and image gen will hollow out a lot of 3d work that isn't explicitly gaming oriented IMO, and eventually even that work may go away once better World-Action-Models come online.
- ThrowawayTestr 1mo agoGood topology is 99% of good 3D models.
- pennomi 1mo agoWe have AI sculpting and animating in the same way that we have AI illustrations - it’s better than the average human but miles and miles away from a professional artist. Most of the feel of a game requires intentional and consistent artistic design, which AI cannot yet replicate, 2D or 3D.
- avaer 1mo agoThe most interesting/surprising development for me has been that LLMs are getting better and better at synthesizing procedural content. Fable 5 can write procedural texture shaders, CSG, surface meshing, l-systems, even rigging + realtime IK, etc. and do it all in a visual loop within an environment. The model understands the topology, and can tune things with mathematical precision where necessary. Which is much closer to using 3D tools like an artist would.