TripoSG

TripoSG is VAST-AI-Research's open model for high-fidelity 3D shape synthesis from images, described in the February 2025 arXiv paper and released with weights and inference code in March 2025. It pairs a 1.5B-parameter rectified-flow transformer with an SDF-based VAE trained on 2 million curated image-SDF pairs, using 2048 latent tokens, with hybrid SDF, surface-normal, and eikonal supervision.

The release includes the rectified-flow model, the VAE, inference code, and an interactive demo, with weights on Hugging Face and a minimum requirement of about 8GB VRAM. It targets sharp geometric detail and semantic consistency across photorealistic images, cartoons, and sketches. Note on dating: the March 2025 release date is era-level from the repo's update notes; verify the license on the model card before commercial use.