ComfyUI-SDCPP

ComfyUI-SDCPP

A custom node for ComfyUI that integrates the powerful stable-diffusion.cpp inference engine. This node brings high-performance, gguf quantized model support to ComfyUI, with massive optimization specifically for Apple Silicon (Metal) devices to efficiently run huge models like Flux!

🌟 Features

• GGUF Support In ComfyUI: Automatically patches ComfyUI's model loaders to recognize .gguf extensions. You can effortlessly load unet, vae, and clip files in GGUF format!
• Apple Silicon / Metal Optimization:
  • Forces GGML to correctly locate the ggml-metal.metal shader files, preventing silent CPU fallbacks.
  • Implements enable_mmap and offload_params_to_cpu for native zero-copy streaming directly from ssd to Unified Memory. Massively reduces swapping out RAM, preventing OOMs (Out Of Memory) entirely.
• Flash Attention: Greatly speeds up generation and reduces VRAM/RAM load footprint.
• ComfyUI Compatibility: Fully integrates with ComfyUI's step progress bar, interruption triggers (Cancel button), and robustly translates the output directly into perfectly contiguous standard IMAGE tensors, seamlessly compatible with downstream ComfyUI PyTorch nodes.

📦 Installation

1. Ensure your Python environment has the stable_diffusion_cpp bindings installed:

   pip install stable-diffusion-cpp-python
   

2. Navigate to your ComfyUI custom_nodes folder and clone this repository (or copy the ComfyUI-SDCPP folder inside):

   cd ComfyUI/custom_nodes
   git clone https://github.com/your-username/ComfyUI-SDCPP.git
   

3. Boot up ComfyUI!

🚀 Usage

Inside ComfyUI:

1. Double click the workspace (or search the node list) and add: SD.cpp Flux (UNet/VAE/CLIP).

2. Place your downloaded .gguf files into the standard ComfyUI directories:

   • ComfyUI/models/unet/
   • ComfyUI/models/clip/
   • ComfyUI/models/vae/ (Node specifically ensures any qwen / mistral / llama named text encoders route to the llm_path for correct Flux handling).

3. Select your UNet, CLIP, and VAE.

4. Set your prompt, resolution, and steps, then click Queue Prompt.

🔧 Acknowledgements

Powered by stable-diffusion.cpp and ggml.