bytedance-seed/depth-anything-3 — reverse-engineered prompt

Build me a local Python tool for Depth Anything 3 that lets me upload one photo, a set of photos, or a video and get useful visual geometry results without needing to understand the research paper.

I want a simple web interface with a gallery where I can preview the original images, depth maps, confidence maps, estimated camera poses when available, and a basic 3D reconstruction. Also include a command line mode so I can batch process folders. Let me choose between the main models, the metric depth model, the monocular model, and the nested model if those weights are available.

The tool should work with images with or without known camera poses, and it should export common results like depth images, npz files, ply files, glb files, and 3D Gaussian videos when supported. Please make setup straightforward, include example commands, and add clear error messages for missing GPU memory or missing optional packages.

Look up the current docs online if you need to.