# Exploring Remote Zoo Models This section walks you through the process of using **remotely-sourced models** from the FiftyOne Model Zoo—models whose definitions are hosted on public GitHub repositories or accessible via external URLs. With this approach, you can: - Seamlessly integrate custom models hosted on GitHub or cloud archives - Reproduce and share models across teams and projects using standardized links - Apply these models to your datasets within FiftyOne just like built-in zoo models FiftyOne’s flexible zoo API supports both built-in and remote models. That means whether you’re pulling a model from `voxel51/openai-clip`, `ultralytics/yolov5`, or your own repository, the workflow remains consistent and intuitive. ## In this notebook, you’ll learn how to: - Specify remote model sources using GitHub repos, branches, or commit references - Load and apply these models to your datasets - Visualize the predictions directly in the FiftyOne App > 💡 To use private GitHub repositories, be sure to set the `GITHUB_TOKEN` environment variable with your personal access token. > 💡 Before starting, ensure you’ve installed the required packages: ### Using Florence2 as Remotely Sourced Zoo Model Original documentation of this Remotely Sourced Zoo Model is [here](https://github.com/harpreetsahota204/florence2) For context, here is the first image: # Setup Zoo Model The three captioning operations require no additional arguments beyond selecting the operation type. Supported `detail_level` values: - `basic` - `detailed` - `more_detailed` To change the caption detail level: The operations for `detection`, `dense_region_caption`, `region_proposal` don’t require additional parameters for general use. However, `open_vocabulary_detection` requires a `text_prompt` parameter to guide the detection towards specific objects. The results are stored as Detections objects containing bounding boxes and labels: Or you can use the caption field: For dense detections, this runs without a prompt and returns all detectable objects. Phrase grounding requires either a direct caption or a reference to a caption field. You can provide this in two ways: When you want to use a Field of a Sample for grounding, you use the following pattern: Segmentation requires either a direct expression or a reference to a field containing expressions. Similar to phrase grounding, you can provide this in two ways: When you want to use a Field of a Sample for grounding, you use the following pattern: Basic OCR (“ocr”) requires no additional parameters and returns text strings. For OCR with region information (`ocr_with_region`), you can set `store_region_info=True` to include bounding boxes for each text region: Remotely-sourced models expand the power and flexibility of the FiftyOne Model Zoo by allowing you to access and deploy models from external GitHub repositories or public URLs. Whether you’re leveraging a popular open-source model or integrating one from your own team, this workflow makes it easy to apply and visualize predictions in your datasets.