Robbyant/lingbot-map ? reverse-engineered prompt

I want to build a tool that can create 3D models of real-world places from a video stream. It should take a video file or a folder of images as input and generate a 3D point cloud map of the scene, ideally processing the footage in real-time at a good frame rate.

The main feature is this live reconstruction. As the video is processed, the 3D map should be built up, and I want to be able to see the reconstruction happen in an interactive viewer, maybe one that runs in the browser. For any outdoor scenes, it's important that the tool can automatically detect and remove the sky to get a cleaner model.

Please set this up as a Python project using PyTorch. It should download the necessary AI models from Hugging Face and be configured to run the included demo script with the example scenes. Use whatever dependencies are recommended for the best performance on a CUDA-enabled GPU.