ComfyUI SAM3 Integration

ComfyUI custom nodes for Meta's SAM 3 (Segment Anything with Concepts).

SAM 3 enables open-vocabulary text-based segmentation - segment objects using natural language prompts like "person", "red car", "shoe" without manual clicking or drawing.

---

Features

• 🎯 Text-based segmentation: Use natural language to describe what to segment
• 🔀 Multi-class support: Comma-separated prompts like "person,car,dog"
• 🎭 Instance merging: All detected instances automatically merged
• 🖼️ Dual outputs: Mask + masked image
• ⚡ GPU acceleration: bf16/fp16/fp32 precision support
• 📦 Auto-download: Model downloads from Hugging Face automatically

---

Installation

Via ComfyUI Manager (Recommended)

Search for ComfyUI-segment-anything-3 in ComfyUI Manager and install.

Manual Installation

cd ComfyUI/custom_nodes/
git clone --recursive https://github.com/wzyfromhust/ComfyUI-segment-anything-3.git
cd ComfyUI-segment-anything-3

Note: Use --recursive to include SAM2 and SAM3 submodules.

---

Nodes

1. DownloadAndLoadSAM3Model

Loads the SAM3 model.

Parameters:

• device: cuda/cpu
• precision: bf16 (default), fp16, fp32
• confidence_threshold: 0.0-1.0 (default: 0.5)

2. Sam3Segmentation

Segments objects using text prompts.

Parameters:

• sam3_model: Model from loader
• image: Input image
• text_prompt:
  • Single: "person"
  • Multiple: "person,shoe,car"
  • Descriptive: "red car", "person wearing blue"
• keep_model_loaded: Keep in VRAM (default: false)

Outputs:

• mask: Segmentation mask [1, H, W]
• image: Masked image [1, H, W, C]

---

Quick Start

LoadImage → Sam3Segmentation → PreviewImage
              ↑
    (Down)Load SAM3 Model

Text prompt examples:

• "person" - All people
• "person,shoe" - People and shoes (merged)
• "red car" - Red cars only
• "person wearing blue" - People in blue clothes

---

Model Setup

Model path: /root/ComfyUI/models/sam3/sam3.pt

Auto-download: If model doesn't exist, it downloads from Hugging Face automatically.

Requirements:

huggingface-cli login
# Accept terms at https://huggingface.co/facebook/sam3

Model specs:

• Size: ~850 MB
• Parameters: 848M

---

Documentation

• 📘 English Documentation
• 📗 中文文档

---

Examples

Single Class

text_prompt = "person"  # Segments all people

Multiple Classes (Auto-merged)

text_prompt = "person,car,dog"  # Segments people, cars, and dogs

Descriptive

text_prompt = "red car"  # Only red cars
text_prompt = "person wearing blue"  # People in blue

---

Technical Details

• Built on SAM3 official implementation
• Strictly aligned with official inference pipeline
• Uses torch.autocast with bf16 for optimal performance
• Supports multi-prompt inference with automatic merging

---

Requirements

• ComfyUI
• PyTorch with CUDA (for GPU acceleration)
• Hugging Face account (for model download)

---

Credits

• SAM3 Official - Meta AI Research
• ComfyUI SAM2 - SAM2 integration reference

---

License

This project integrates:

• SAM3: Apache 2.0 License
• SAM2: Apache 2.0 License

---

Links

• 📄 SAM3 Paper
• 🤗 SAM3 Model
• 💻 GitHub Repository