ComfyUI-SAM3

ComfyUI integration for Meta's SAM3 (Segment Anything Model 3) - enabling open-vocabulary image and video segmentation using natural language text prompts.

https://github.com/user-attachments/assets/323df482-1f05-4c69-8681-9bfb4073f766

Installation

Install via ComfyUI Manager or clone to ComfyUI/custom_nodes/:

cd ComfyUI/custom_nodes/
git clone https://github.com/PozzettiAndrea/ComfyUI-SAM3.git
cd ComfyUI-SAM3
python install.py

Optional: GPU Acceleration for Video Tracking

For 5-10x faster video tracking, install GPU-accelerated CUDA extensions:

python speedup.py        # Auto-detects your GPU, ~3-5 min compilation

This is optional and only benefits video tracking performance. Image segmentation works fine without it. The script will:

• Auto-detect your GPU architecture and compile only for your specific GPU (75-80% faster than previous versions)
• Auto-install CUDA toolkit via conda/micromamba if needed
• Compile GPU-accelerated extensions (torch_generic_nms, cc_torch)

Requirements: NVIDIA GPU with compute capability 7.5+ (RTX 2000 series or newer), conda/micromamba environment recommended.

RTX 50-series (Blackwell): Experimental support available via python speedup_blackwell.py (~45-60 sec compilation). May compile successfully but runtime stability not guaranteed due to PyTorch lacking official sm_120 support. Falls back to CPU mode if compilation fails. Track PyTorch support at pytorch/pytorch#159207.

Troubleshooting

SAM3 nodes not appearing in ComfyUI

If SAM3 doesn't load and you see "running in pytest mode - skipping initialization" in the logs, this is a false positive detection.

Solution: Set the environment variable before starting ComfyUI:

# Linux/Mac
export SAM3_FORCE_INIT=1

# Windows
set SAM3_FORCE_INIT=1

This forces SAM3 to initialize even if pytest is detected in your environment.

Examples

Nodes

Image Segmentation

• LoadSAM3Model - Load SAM3 model for image segmentation
• SAM3Segmentation - Segment objects using text prompts ("person", "cat in red", etc.)
• SAM3CreateBox - Create bounding box prompts (normalized coordinates)
• SAM3CreatePoint - Create point prompts with positive/negative labels
• SAM3CombineBoxes - Combine multiple box prompts
• SAM3CombinePoints - Combine multiple point prompts

Video Tracking

• SAM3VideoModelLoader - Load SAM3 model for video tracking
• SAM3InitVideoSession - Initialize video tracking session
• SAM3InitVideoSessionAdvanced - Advanced session initialization with custom settings
• SAM3AddVideoPrompt - Add object prompts to track in video
• SAM3PropagateVideo - Propagate object tracking through video frames

Interactive Tools

• SAM3PointCollector - Interactive UI for collecting point prompts
• SAM3BBoxCollector - Interactive UI for drawing bounding boxes

---

Quick Start

1. Add LoadSAM3Model node (first run downloads ~3.2GB model from HuggingFace)
2. Add SAM3Segmentation node, connect model
3. Enter text prompt: "person", "cat in red", "car on the left"
4. Get masks, visualization, boxes, and confidence scores

Text Prompt Examples:

• "shoe", "cat", "person" - Single objects
• "person in red", "black car" - With attributes
• "person on the left", "car in background" - Spatial relations

Video Tracking:

1. Use SAM3VideoModelLoader instead of LoadSAM3Model
2. Initialize session with SAM3InitVideoSession
3. Add prompts with SAM3AddVideoPrompt
4. Propagate with SAM3PropagateVideo

Credits

• SAM3: Meta AI Research (https://github.com/facebookresearch/sam3)
• ComfyUI Integration: ComfyUI-SAM3
• Interactive Points Editor: Adapted from ComfyUI-KJNodes by kijai (Apache 2.0 License). The SAM3PointsEditor node is based on the PointsEditor implementation from KJNodes, simplified for SAM3-specific point-based segmentation.