This repository offers a custom node for ComfyUI that implements YOLO-NAS-POSE, utilizing TensorRT to achieve rapid pose estimation at over 100 frames per second. It is specifically designed to integrate with the openpose controlnet in an experimental capacity.
- Enables ultra-fast pose estimation, enhancing real-time applications.
- Supports various model sizes and precision settings for tailored performance.
- Facilitates easy engine building and usage through a user-friendly interface.
Context
This tool is a custom node for ComfyUI that allows users to perform pose estimation using the YOLO-NAS-POSE model, optimized with TensorRT for high-speed processing. Its purpose is to provide efficient pose detection capabilities within the ComfyUI framework, making it suitable for applications requiring real-time analysis.
Key Features & Benefits
The primary functionality of this tool includes the ability to operate with multiple model sizes (small, medium, and large) and precision levels (FP32 and FP16), allowing users to choose the configuration that best fits their hardware and performance needs. This flexibility is crucial for optimizing resource usage and achieving desired frame rates based on the specific application.
Advanced Functionalities
The tool includes two methods for building TensorRT engines: an automated EngineBuilder node that simplifies the process for users and a manual method for those who prefer more control. This dual approach caters to both novice users and advanced practitioners, enabling them to create a custom setup that aligns with their workflow preferences.
Practical Benefits
By integrating this custom node into ComfyUI, users can significantly enhance their workflow efficiency, gaining control over pose estimation tasks while maintaining high-quality outputs. The ability to achieve over 100 FPS allows for real-time applications, making this tool particularly valuable for projects in domains such as gaming, animation, and interactive media.
Credits/Acknowledgments
The development of this tool is credited to the creators of the YOLO-NAS-POSE model, specifically the super-gradients team, who provided the foundational technology upon which this implementation is built. The project is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license, allowing for free access and modification under specified conditions.