ComfyUI Extension: ComfyUI-PixArt_XL

Authored by ShmuelRonen

Created

Updated

2 stars

A ComfyUI extension that integrates PixArt-Alpha models directly into ComfyUI with advanced memory management.

Custom Nodes (0)

    README

    ComfyUI-PixArt_XL

    Support My Work

    If you find this project helpful, consider buying me a coffee:

    Buy Me A Coffee

    A ComfyUI extension that integrates PixArt-Alpha models directly into ComfyUI with advanced memory management.

    image

    Features

    • Use PixArt-Alpha XL models directly in ComfyUI
    • Smart memory management using both GPU and system RAM
    • Automatic fallback to lower resolutions with high-quality upscaling
    • Support for both DPM-Solver and SA-Solver schedulers
    • Full precision processing for maximum stability

    Installation

    1. Clone this repository into your ComfyUI custom nodes folder:
    cd ComfyUI/custom_nodes
    git clone https://github.com/ShmuelRonen/ComfyUI-PixArt_XL.git
    
    1. Install dependencies:
    pip install sentencepiece
    
    1. Restart ComfyUI

    Usage

    After installation, you'll have access to two new nodes:

    PixArtAlpha ModelLoader

    Automatic loads a PixArt-Alpha model from Hugging Face.

    Inputs:

    • base_model_path: Path to the model (default: "PixArt-alpha/PixArt-XL-2-1024-MS")

    Outputs:

    • model: The loaded model

    PixArtAlpha Generation

    Generates images using the PixArt-Alpha model.

    Inputs:

    • model: The PixArt-Alpha model
    • positive: Positive prompt
    • negative: Negative prompt
    • width, height: Image dimensions
    • steps: Number of sampling steps
    • guidance_scale: How closely to follow the prompt
    • schedule: Sampling scheduler
    • seed: Random seed

    Outputs:

    • image: The generated image

    Memory Optimization

    This implementation includes several advanced techniques to maximize memory efficiency:

    • CPU component offloading
    • Attention slicing
    • VAE optimization
    • Periodic CUDA cache clearing
    • Resolution fallback with quality-preserving upscaling

    Acknowledgments