facebookresearch/sam2 ? reverse-engineered prompt

Build me a simple app and notebook setup for Meta SAM 2 that lets me segment things in images and videos with a few clicks.

I want to upload an image or short video, click on the object I care about, and have the model draw the mask and keep tracking it through the video. It should also support automatic mask generation for images when I do not want to give prompts. Please include a clean demo page or notebook flow that shows how to load a checkpoint, run inference, and visualize the results clearly.

Make it easy to swap between the small and larger checkpoints, and keep the setup instructions straightforward so someone can get it running locally on a GPU machine. If there are current install or usage details I need, look them up online and use the latest recommended approach.