InstantMesh

InstantMesh generates 3D meshes from a single image. It's for creators working with image-based 3D assets and researchers studying 3D reconstruction.

• Local interface: A Gradio demo lets you generate meshes on your own machine, with Docker support and a command-line alternative.
• Background removal: The tool uses rembg to separate the foreground object; images with an existing alpha mask can skip this step.
• Reconstruction models: You can choose among sparse-view reconstruction variants based on LRM/Instant3D, with instant-mesh-large as the default and a customized Zero123++ model for white-background image generation.
• Research training: The project includes reconstruction training code and Zero123++ fine-tuning code, but doesn't include the training dataset.

InstantMesh uses Python and PyTorch with CUDA for GPU processing. The local demo can run on a single GPU or use two GPUs to reduce memory use. The inference script can download the required models automatically. The project is open source under Apache 2.0.