ComfyUI EdgeTAM

A ComfyUI custom node implementation of EdgeTAM (On-Device Track Anything Model) for efficient, interactive video object tracking.

Overview

EdgeTAM is an optimized variant of SAM 2 designed for on-device execution, running 22× faster than SAM 2 while maintaining high accuracy. This custom node provides a seamless and interactive workflow for video object segmentation and tracking within ComfyUI.

Features

• Interactive Video Object Tracking: Pause the workflow at any time to draw a mask on the first frame of your video. The editor provides a live preview of the segmentation.
• Automated Workflow Support: For batch processing, you can bypass the interactive editor by providing a JSON string with pre-defined mask points.
• High Performance: Optimized for real-time inference on consumer hardware.
• Automatic Installation: The required EdgeTAM library and model checkpoint are automatically installed on the first run.

Installation

1. Clone this repository into your ComfyUI custom_nodes directory:

   cd ComfyUI/custom_nodes
   git clone https://github.com/your-repo/comfyui_EdgeTAM.git
   

2. Restart ComfyUI. The necessary dependencies and the EdgeTAM model will be installed automatically the first time you run a workflow.

Usage

This package provides two main nodes for a complete interactive tracking workflow.

1. Interactive Mask Editor

This node is the core of the interactive workflow.

• Inputs:
  • image: Connect the video frames (as an IMAGE batch) from a loader node.
  • optional_mask_data (string): Leave this disconnected for interactive use. For automation, you can connect a string node containing a JSON object.
• Behavior:
  • Interactive Mode: When the workflow runs with the optional_mask_data input disconnected, it will pause and open a full-screen editor. Here you can:
    • Left-click to add a positive ("include") point.
    • Right-click to add a negative ("exclude") point.
    • Use the Preview Mask button to see the segmentation result in real-time.
    • Click Save and Continue to send the mask data to the next node.
  • Automation Mode: If you provide a valid JSON string to the optional_mask_data input, the editor is skipped entirely. This is ideal for batch processing or when you want to programmatically define the mask. The JSON must have the following structure:

    {
      "points": [[x1, y1], [x2, y2], ...],
      "labels": [1, 0, ...]
    }
    

    • points: A list of [x, y] coordinate pairs.
    • labels: A corresponding list where 1 is an include point and 0 is an exclude point.

2. EdgeTAM Video Tracker

This node performs the actual video tracking.

• Inputs:
  • video_frames: Connect the same video frames from your loader node.
  • mask_data: Connect the mask_data output from the InteractiveMaskEditor node.
• Output:
  • tracked_frames: The original video frames.
  • masks: The generated segmentation masks for each frame.
  • overlay_frames: The original frames with the masks drawn on top.

Example Workflow

graph TD
    A[Load Video] --> B(InteractiveMaskEditor);
    A --> C{EdgeTAMVideoTracker};
    B -->|mask_data| C;
    C --> D[Preview Image];

    style B fill:#e67e22,stroke:#333,stroke-width:2px

Requirements

• Python >= 3.10
• PyTorch >= 2.3.1
• ComfyUI

License

This project follows the EdgeTAM license (Apache 2.0). See the LICENSE file for details.

Credits

Based on EdgeTAM by Meta Reality Labs:

• Paper: "EdgeTAM: On-Device Track Anything Model" (CVPR 2025)
• Repository: https://github.com/facebookresearch/EdgeTAM