GLOSSARY
Text-to-3D
Text-to-3D is a class of AI systems that take a description in plain English and return a 3D model — usually a mesh, sometimes with textures.
Definition
The first generation of text-to-3D (DreamFusion, Magic3D, around 2022–2023) optimized a NeRF or similar neural representation against a 2D image diffusion model — the "score distillation" approach. Output was soft, view-inconsistent, and hard to print.
The current generation (2024–) generates intermediate 2D views of the object, then reconstructs an explicit polygon mesh from those views. This is the approach Automatic3D uses: generate a consistent multi-view collage, hand it to a mesh reconstruction model, output STL.
Why it matters
Text-to-3D fills a gap CAD does not: getting a recognizable shape from a description without modelling it. For visual objects — figurines, props, decorative items, organic shapes — it is faster than sculpting and more flexible than searching Thingiverse.
Common confusion
Text-to-3D is not a CAD replacement. You cannot specify dimensions, tolerances, or mating features in a prompt. For functional mechanical parts, you still want CAD. Text-to-3D is for the cases where "a thing that looks like X" is the entire spec.
Most text-to-3D tools target digital content (game assets, AR); a smaller subset targets 3D printing. The output meshes differ in important ways — printability vs render quality, manifold guarantees vs textured fidelity.