🤗 Hugging Face | 🤖 ModelScope | 💬 WeChat (微信) | 🎮 Discord

中文说明

ComfyUI plugin for QuantFunc — the fastest diffusion model inference engine. Run quantized text-to-image and image editing models at 2x–11x speed with zero Python model dependencies.

Key features:

• Native C++/CUDA acceleration via libquantfunc.so / quantfunc.dll
• SVDQ (offline quantization) + Lighting (runtime quantization) dual engine
• Zero-cost LoRA stacking
• Image editing with reference images
• Export runtime-quantized models with LoRA fusion support
• Auto-update from ModelScope

🎯 Ease of Use

• v2 loaders — separate MODEL / CLIP / VAE sockets feed a Build Pipeline node, so models wire up the ComfyUI-native way instead of one monolithic loader.
• Universal format adapters — load diffusers / BFL (Flux) / nunchaku SVDQ / bundled-checkpoint / HF layouts automatically, with no manual conversion.
• Base Model Auto Loader with one-click download; the plugin also auto-pulls the matching engine on first startup.

🧩 Model Support

• SVDQ (offline quantization) + Lighting (runtime BF16/FP16 → 4-bit) dual engine.
• Pipelines: Z-Image · QwenImage · QwenImage-Edit · Flux.2 Klein.
• Full GPU coverage (engine 0.0.07): consumer RTX 20 / 30 / 40 / 50-series, datacenter A100 / H100 / H200 / B100 / B200 / GB300, workstation RTX 6000 Ada / RTX PRO 6000 Blackwell — across CUDA 12 & 13.

⚡ Performance

• Consumer GPUs run native SASS — no first-run JIT compile stall on 20/30/40/50-series (datacenter/workstation cards JIT once, then cache).
• Native FP4 (NVFP4) on Blackwell (SM120) — the fastest 4-bit path.
• QFRAW raw staging for reference images & masks skips the PNG/BMP encode (~80 ms saved per ref).
• Multi-pipeline CPU↔GPU coexistence — swap pipelines without a full reload; idle workers auto-free VRAM.

✨ New Features

• Inpainting — MASK input plus Mask Config and Mask Scale By nodes (white = regenerate, black = preserve), mirroring ComfyUI's SetLatentNoiseMask.
• Build Pipeline node (v2 assembly) with per-component precision control.
• Robust worker-process architecture — CPU↔GPU model swap + zombie-worker cleanup.

🛡️ Stability & Security

• Fixed a /dev/shm RAM leak — edit/inpaint staging files are now always cleaned up.
• Zip-slip guard on dependency-archive extraction.
• IPC bound-check on the worker → host image transfer.

The plugin auto-pulls the matching engine on startup: bumping comfy to 0.0.02 lets the updater fetch engine 0.0.07 from ModelScope (older comfy stays capped at engine 0.0.06).

cd ComfyUI/custom_nodes
git clone https://github.com/QuantFunc/ComfyUI-QuantFunc.git

The plugin will automatically download the latest compatible libquantfunc.so (Linux) or quantfunc.dll (Windows) from ModelScope on first startup. No manual binary download needed.

1. Download or clone this repository into ComfyUI/custom_nodes/:

ComfyUI/
└── custom_nodes/
 └── ComfyUI-QuantFunc/
 ├── __init__.py
 ├── nodes.py
 ├── worker.py
 ├── auto_update.py
 └── bin/
 ├── linux/
 │ └── version.json
 └── windows/
 └── version.json

2. Start ComfyUI — the plugin auto-downloads the library binary on first run.

3. (Optional) To skip auto-download, manually place the binary:

   • Linux: Download libquantfunc.so → bin/linux/
   • Windows: Download quantfunc.dll → bin/windows/

# CUDA 12 runtime libraries
sudo apt install cuda-libraries-12-8
# or individual packages:
sudo apt install libcublas-12-8 libcurand-12-8 libcusolver-12-8 libcusparse-12-8 libnvjitlink-12-8
# cuDNN 9
sudo apt install libcudnn9-cuda-12
# --- OR ---
# CUDA 13 runtime libraries
sudo apt install cuda-libraries-13-0
# or individual packages:
sudo apt install libcublas-13-0 libcurand-13-0 libcusolver-13-0 libcusparse-13-0 libnvjitlink-13-0
# cuDNN 9
sudo apt install libcudnn9-cuda-13

• NVIDIA Driver ≥ 560 (provides CUDA runtime DLLs)
• Visual C++ Redistributable 2015-2022 (download)
• cuDNN 9.x (download)

Auto-update requires modelscope Python package:

pip install modelscope

If modelscope is not installed, auto-update is silently skipped. You can manually download binaries from:

• https://www.modelscope.cn/models/QuantFunc/Plugin

After starting ComfyUI, check the console for:

[QuantFunc] Checking for updates (plugin v0.0.01, lib v0.0.01)...
[QuantFunc] Library is up to date (v0.0.01)

If the library was not found:

[QuantFunc] No library found, checking ModelScope for download (plugin v0.0.01)...
[QuantFunc] Downloading libquantfunc.so v0.0.01 from ModelScope...
[QuantFunc] Updated libquantfunc.so to v0.0.01. Restart ComfyUI to use the new version.

See doc/ for detailed tutorials and workflow_sample/README.md for node reference.

The easiest way to get started — import the Easy Gen workflow, pick a model from the dropdown, and the plugin auto-downloads everything. No manual model downloads or path configuration needed.

Beginners: Easy Gen →

The Lighting backend provides runtime quantization — it uses the Lighting engine to quantize any diffusers-format BF16/FP16 model (e.g., Qwen/Qwen-Image-Edit-2511) to 4bit at load time for accelerated inference. Just set model_backend to lighting and leave transformer_path empty — no pre-quantized model download needed.

Tutorial 1: Runtime Quantization →

The Lighting export saves all runtime-quantized models to disk, so you don't need to re-quantize on every startup. If you've also stacked LoRAs, they are permanently fused into the exported weights — no LoRA nodes needed, no re-quantization, load and go.

Tutorial 2: Export Runtime-Quantized Models →

QuantFunc has pre-exported commonly used models (runtime-quantized and ready to use). Download them directly from ModelScope or HuggingFace — same 2x–11x inference speedup as runtime quantization, but with faster loading since the quantization step is skipped.

Tutorial 3: Download & Use Pre-exported Models →

Import from workflow_sample/:

See QuantFunc Plugin License.

Join our community for support, updates, and discussions:

• 🎮 Discord server
• 💬 Scan the QR code below to join our WeChat group: