TripoSG

A 1.5B parameter rectified flow transformer sits at the core of TripoSG, VAST AI Research's image-to-3D foundation model, paired with a VAE that represents geometry as signed distance functions and was trained with hybrid supervision on 2 million curated image-SDF pairs. The linear-trajectory flow formulation generates high-fidelity meshes from photorealistic images, cartoons, and sketches alike, and a March 2025 companion release, TripoSG-scribble, adds a 512-token model for generation from scribbles plus text prompts. Weights download automatically from Hugging Face on first run, installation is the standard conda plus pip requirements routine on Python 3.10, and inference needs a CUDA GPU with at least 8 GB of VRAM. Both code and released weights carry the MIT license, so commercial use is permitted, which distinguishes it from most open 3D generators that ship restrictive model licenses. The team behind the commercial Tripo service maintains the project, it holds about 1.7k GitHub stars, and it serves as the open research counterpart to the company's hosted generation pipeline.