facebookresearch/sam3 ? reverse-engineered prompt

Build me a Python library for Meta's SAM 3 model that lets anyone segment and track objects in images or videos using simple text descriptions.

You type a phrase like the dog on the left or player in red jersey and it finds every matching instance, draws precise masks, returns boxes and confidence scores, even when the scene has many similar looking things. It should also accept clicks, boxes or example patches as extra prompts.

Support both one off photos and full video files with automatic tracking across frames. Keep the interface really straightforward, something like load the model, feed it an image or video path plus a text prompt, and get the results back.

Include a few ready to run examples that show image segmentation, video tracking, and interactive refinement. Use the latest model weights from Hugging Face and make installation work with recent PyTorch and CUDA. If you need the exact architecture details just pull the paper.