InstantMesh

InstantMesh generates a 3D mesh from a single image on your own machine. It's for creators exploring image-based 3D assets and researchers working on 3D reconstruction. The Python project is open source under Apache 2.0 and uses PyTorch with CUDA for GPU processing.

You can use a local Gradio interface or process images through the command line. Docker is also supported. The demo can run on one GPU or split work across two GPUs to reduce memory use. A hosted Hugging Face demo offers a separate way to try it; the local application runs on your hardware.

The default output is an OBJ mesh with vertex colors. You can also export a mesh with a texture map, though that takes longer. Foreground segmentation uses rembg to separate the object from its background, and images that already contain an alpha mask can bypass that step. The tool can also save a video of the generated result.

Its reconstruction models build on the LRM/Instant3D architecture, with mesh and NeRF variants available. It includes a customized Zero123++ model for generating views with white backgrounds, and the inference script can download the required model weights automatically.

For research work, the project provides reconstruction training code and Zero123++ fine-tuning code alongside the weights. The training dataset isn't included. Apache-2.0 describes the project code; downloaded reconstruction, Zero123++ and background-removal components have their own model terms, which should be checked separately.