Summary

QuantFunc/ComfyUI-QuantFunc accelerates diffusion model inference within ComfyUI using a native C++/CUDA engine. It targets users seeking significant speedups (2x-11x) for text-to-image and image editing, eliminating Python dependencies during runtime and enabling efficient quantized model usage.

How It Works

The core is the native C++/CUDA libquantfunc.so/quantfunc.dll engine. It employs a dual quantization strategy: SVDQ (offline) and Lighting (runtime BF16/FP16 → 4-bit). This, combined with zero-cost LoRA stacking/fusion, drastically reduces inference latency. Universal model format adapters and a robust worker-process architecture enhance efficiency.

Quick Start & Requirements

• Install via git clone https://github.com/QuantFunc/ComfyUI-QuantFunc.git into ComfyUI/custom_nodes. The plugin auto-downloads the engine binary on first startup.
• Prerequisites: NVIDIA RTX 20+ GPU (CC 7.5+), 8 GB VRAM, NVIDIA Driver ≥ 560, CUDA 13.0+, cuDNN 9.x, Linux (glibc 2.31+) or Windows 10/11.
• Dependencies: pip install modelscope recommended for auto-update.
• Links: GitHub repository (implied), ModelScope.

Highlighted Details

• Achieves 2x–11x inference speedup.
• Zero Python model dependencies during inference.
• Dual quantization: SVDQ (offline) and Lighting (runtime BF16/FP16 → 4-bit).
• Zero-cost LoRA stacking and permanent fusion.
• Universal adapters load various model types (diffusers, Flux, SVDQ, HF) automatically.
• Full GPU coverage (consumer, datacenter, workstation) across CUDA 12/13.
• Native FP4 inference on Blackwell GPUs.
• Includes inpainting capabilities.

Maintenance & Community

• Features an auto-update mechanism for the engine binary from ModelScope.
• Community support available via Discord and WeChat.

Licensing & Compatibility

• Specific license not explicitly stated; README refers to "QuantFunc Plugin License".
• Primarily compatible with ComfyUI. Commercial use requires clarification of the underlying license.

Limitations & Caveats

• Strict NVIDIA hardware (RTX 20+) and driver/CUDA/cuDNN version requirements.
• Auto-update relies on the modelscope Python package.
• Unclear license terms in README may impact commercial adoption.