SAM3D Body with Rerun (github.com)

An unofficial, one‑command playground combines Meta’s new SAM3D Body (DINOv3) and promptable SAM3 into an interactive demo that streams 3D inspection via Rerun and serves a user interface in Gradio. The repo ties together SAM3 (promptable segmentation), SAM3D Body (3D human reconstruction), and a relative‑depth predictor (MogeV1Predictor/monopriors) so you can upload an RGB image, get promptable SAM masks, toggle “Log relative depth” to stream predicted depth, and inspect the reconstructed 3D outputs live in the embedded Rerun viewer. Setup uses the Pixi package manager for a one‑command launch (git clone …; pixi run app) on Linux with an NVIDIA GPU. Technically important: the SAM3 and SAM3D Body checkpoints are gated on Hugging Face—request access and authenticate (HF_TOKEN or huggingface‑cli login) before the demo’s first download. First run pulls SAM3, SAM3D Body, and the relative‑depth model. The repo also provides a dev CLI/shell for batch runs, a mode that returns SAM3 masks only, and example data. The code is dual‑licensed (Apache‑2.0 and MIT); upstream model assets keep their original licenses. This demo is a useful, practical tooling integration for researchers and engineers wanting fast interactive exploration of promptable segmentation tied to 3D body reconstruction and depth streaming.