ComfyUI_SAM2UltraV2

ComfyUI SAM2 Ultra V2

A fully modular and high-performance implementation of SAM2 (Segment Anything Model 2) for ComfyUI.

Features

• 🚀 High Performance: Model loading optimization with 50-80% faster repeated inference
• 🎯 Dual Prompt Support: Combine bbox and keypoint prompts for enhanced accuracy
• ⭕ Negative Point Sampling: Ring-region negative sampling for better foreground/background separation
• 👤 Upper Body Segmentation: Hip-based intelligent cutting for portrait segmentation
• 📦 Fully Modular: Clean code architecture with separated concerns
• 🔧 Independent Package: No dependencies on other custom nodes

Nodes

1. SAM2 Ultra V2: Load Model

Load and cache SAM2 model for reuse across multiple segmentations.

Inputs:

• model: SAM2 model selection
• precision: fp16/bf16/fp32
• device: cuda/cpu
• segmentor: single_image/video

Outputs:

• sam2_model: Reusable model object

2. SAM2 Ultra V2: Upper Body

Segment upper body using hip keypoints for intelligent cutting.

Inputs:

• sam2_model: From Load Model node
• image: Input image
• pose_keypoint: OpenPose keypoints
• keypoint_types: upper_body/face/hands
• Detail processing parameters

Outputs:

• image: RGBA image with transparency
• mask: Binary mask
• context_image: Visualization with keypoints and cut line

3. SAM2 Ultra V2: Crop

Bbox-based segmentation with dual prompts and negative sampling.

Inputs:

• sam2_model: From Load Model node
• image: Input image
• bboxes: Detection bounding boxes
• pose_keypoint (optional): For dual-prompt mode
• keypoint_types: none/face/hands
• enable_negative_points: Enable ring sampling
• negative_sample_count: Number of negative points (default: 4)
• negative_margin: Buffer zone size (default: 0.1)
• Detail processing parameters

Outputs:

• image: RGBA image with transparency
• mask: Binary mask
• context_image: Visualization with bboxes and keypoints

Installation

1. Clone this repository into ComfyUI/custom_nodes/:

cd ComfyUI/custom_nodes/
git clone https://github.com/yourusername/ComfyUI_SAM2UltraV2.git

2. Install dependencies:

cd ComfyUI_SAM2UltraV2
pip install -r requirements.txt

3. Restart ComfyUI

Model Download

SAM2 models will be automatically downloaded from Hugging Face on first use. Models are stored in ComfyUI/models/sam2/.

Available models:

• sam2_hiera_tiny.safetensors (Fastest)
• sam2_hiera_small.safetensors
• sam2_hiera_base_plus.safetensors (Recommended)
• sam2_hiera_large.safetensors (Best quality)
• SAM 2.1 versions with -fp16 suffix

Usage Example

Basic Upper Body Segmentation

OpenPose Estimator → Pose Keypoint
                  ↓
SAM2 Ultra V2: Load Model → SAM2 Ultra V2: Upper Body → Output
                         ↓
                    Image Input

Crop Segmentation with Dual Prompts

YOLO Detector → Bboxes
             ↓
SAM2 Ultra V2: Load Model → SAM2 Ultra V2: Crop → Output
                         ↓              ↓
                    Image Input    Pose Keypoint (optional)

Technical Details

Negative Point Sampling

The crop node uses ring-region sampling to place negative points between the bbox and expanded crop region. This helps SAM2 distinguish between:

• Inner region (bbox + margin): Foreground (positive points)
• Ring region: Background (negative points)
• Outer region (crop): Context for segmentation

Module Structure

ComfyUI_SAM2UltraV2/
├── __init__.py              # Node registration
├── config.py                # Constants and configuration
├── model_loader.py          # SAM2 model loading
├── keypoint_utils.py        # Keypoint processing
├── sampling.py              # Negative point sampling
├── imagefunc.py             # Image utilities
├── blendmodes.py            # Blend mode utilities
├── sam2/                    # SAM2 model architecture
└── nodes/
    ├── load_model.py        # Load Model node
    ├── upper_body.py        # Upper Body node
    └── crop.py              # Crop node

Requirements

• Python 3.8+
• PyTorch 1.13+
• ComfyUI
• See requirements.txt for full list

License

MIT License

Credits

• SAM2 Model: Meta AI
• Original SAM2 Ultra implementation: LayerStyle
• Modular refactoring and negative sampling: dz

Changelog

v2.0.0 (2025-11-13)

• Initial modular release
• Added negative point sampling
• Separated into independent package
• Improved code maintainability (38% reduction in complexity)
• Added comprehensive documentation

Support

For issues and feature requests, please open an issue on GitHub.