Text-to-3D

Text-to-3D generates 3D assets — meshes, NeRFs, or Gaussian splats — from text descriptions alone, a capability that barely existed before DreamFusion (2022) showed score distillation sampling could lift 2D diffusion priors into 3D. The field moves at breakneck speed: Magic3D added coarse-to-fine generation, Instant3D achieved single-shot inference, and Meshy and Tripo brought commercial quality. Multi-view consistency remains the core challenge — the "Janus problem" where different viewpoints produce contradictory details. The promise of democratizing 3D content creation for games, VR, and e-commerce is driving massive investment.