naturalpad-studio/instanthmr ? reverse-engineered prompt

Build me a Python app that can take a photo, video, or live camera feed and detect each person, then estimate their full 3D body pose and mesh in real time. I want it to show the 2D skeleton, the 3D body, and basic latency info in a viewer so I can see how fast it’s running. It should work on a single image too, and let me choose between different model sizes for speed vs accuracy, with an option to run a small pose refinement step if that helps the results.

Please make it easy to run from the command line, and also simple to use from another Python script. If there are any model files that need to be downloaded the first time, handle that automatically. Support NVIDIA GPU use when available, but don’t break if I only have CPU. If there’s a good way to smooth motion in videos, include that too. Look up current docs online if you need to.