InstantMesh

Tencent ARC released InstantMesh in April 2024 as one of the first permissively licensed pipelines that could turn a single photo into a usable 3D mesh in seconds. It chains a fine-tuned Zero123++ model, which imagines six consistent views of the object, with a sparse-view large reconstruction model that fuses them into geometry, and it shipped NeRF and mesh variants, training code, a command line script and a Gradio app. Work stopped early: the last code merge came in June 2024, the final commit in January 2025 only edited install notes, and more than a hundred open issues, including recent install failures, have no maintainer reply. It still runs on Linux with an NVIDIA GPU if you recreate the pinned stack of PyTorch 2.1.0 with CUDA 12.1, xformers 0.0.22 and Python 3.10, by hand or with the bundled Docker files; newer driver and library combinations tend to break the build. The code and model card are Apache-2.0, which keeps it attractive for commercial experiments. For new work, Microsoft's TRELLIS or TRELLIS.2 (both MIT) and Tencent's own Hunyuan3D 2.x line produce sharper geometry and proper textures.