ComfyUI SAM3

ComfyUI custom node pack for SAM3 (Segment Anything Model 3) - Meta's state-of-the-art image segmentation model. This extension enables text-prompt-based object segmentation directly within your ComfyUI workflows.

Overview

SAM3 is a powerful zero-shot segmentation model that can identify and segment objects in images using natural language prompts. This custom node pack brings SAM3's capabilities to ComfyUI, allowing you to:

• Segment objects using text descriptions (e.g., "person", "car", "dog")
• Filter results by confidence threshold
• Control minimum object dimensions
• Output individual masks for each detected object
• Generate a combined mask of all detections
• Visualize results with colored overlays, bounding boxes, and confidence scores

Quickstart

1. Clone this repository under ComfyUI/custom_nodes.

2. Install the dependencies:

   pip install -r requirements.txt

3. Model Setup - Choose one of the following options:

   Option A: Auto-download from HuggingFace (recommended)

   • Request model access at https://huggingface.co/facebook/sam3
   • Login to huggingface using hf auth login
   • The model will automatically download on first use

   Option B: Manual checkpoint placement

   • Download sam3.pt manually from https://huggingface.co/facebook/sam3/tree/main
   • Place the checkpoint file at: ComfyUI/models/sam3/sam3.pt
   • The node will automatically detect and use the local checkpoint

4. Restart ComfyUI.

5. Load example workflow from workflow_example/Workflow_SAM3_image_text.json

Features

SAM3 Segmentation Node

Inputs:

• image - Input image to segment (or batch of images for video model)
• prompt (STRING) - Text description of objects to segment (e.g., "person", "car", "building")
• threshold (FLOAT, 0.0-1.0) - Minimum confidence score threshold for detections (default: 0.5)
• min_width_pixels (INT) - Minimum bounding box width in pixels (default: 0)
• min_height_pixels (INT) - Minimum bounding box height in pixels (default: 0)
• use_video_model (BOOLEAN) - Enable video model for temporal tracking across frames (default: False)
• object_ids (STRING, optional) - Comma-separated list of object IDs to track (video model only, e.g., "0,1,2")

Outputs:

• segmented_image (IMAGE) - Visualization with colored mask overlays, bounding boxes, and confidence scores
• masks (MASK) - Batch of individual binary masks, one for each detected object [B, H, W]
• mask_combined (MASK) - Single merged mask containing all detected objects [1, H, W]

Model Modes

Image Model (default)

• Processes each frame independently
• Faster inference
• No temporal consistency between frames
• Best for single images or when frame-to-frame tracking is not needed

Video Model

• Enables temporal tracking across multiple frames
• Assigns consistent object IDs across frames
• Tracks object movement and maintains identity
• Perfect for video sequences or animation frames
• Supports selective tracking via object_ids parameter
• Example: Set object_ids="0,2" to track only objects with IDs 0 and 2

Video Model Features:

• Object IDs are displayed on visualization with format "ID:X score"
• Objects maintain the same ID and color across frames
• Can filter specific objects by providing comma-separated IDs
• Leave object_ids empty to track all detected objects

Example Use Cases:

• Remove backgrounds by segmenting people or objects
• Isolate specific elements in a scene for further processing
• Create masks for inpainting workflows
• Generate batch masks for multiple objects of the same type
• Filter detections by size to focus on foreground/background objects
• Track objects across video frames with consistent IDs (video model)
• Follow specific objects through animation sequences (video model)