3D Model Generation AI Tools
Open-source models and tools for generating 3D meshes, point clouds, and Gaussian splats from text or images.
Producing usable 3D geometry is the slow step in game art, VFX, and robotics simulation. The projects here shortcut it by turning photos, one image, or a text prompt into a mesh or a radiance field. Two camps dominate in 2026. Reconstruction starts from real captures: Meshroom runs classical photogrammetry, Nerfstudio and gsplat train radiance fields from image sets, and feed-forward models like DUSt3R and VGGT regress geometry from unposed images. Generation starts from a prompt or a single view, using multi-view diffusion or large reconstruction transformers to invent what was never observed, as TripoSR, InstantMesh, TRELLIS, and Hunyuan3D-2 do. The tradeoff is accuracy against coverage: reconstruction stays faithful to what the camera saw and says nothing about the unseen side, while generation always returns a complete asset that may not match anything real. Where to start depends on the input. For capture work, Nerfstudio is the gentlest on-ramp, wrapping training, viewing, and export behind one CLI. For single-image generation, Stable Fast 3D and TripoSR each return a textured mesh in roughly a second, which makes them sane baselines before reaching for slower options like TRELLIS. The trap worth checking before quality is licensing. Code and weights are routinely released under different terms here, so an Apache or MIT repository can still ship checkpoints under a non-commercial research license or a vendor community license carrying revenue thresholds and regional carve-outs. That decides whether an asset can ship in a product, and it usually surfaces after the pipeline works.