Nunchaku runs FLUX.1 Dev as a 4-bit SVDQuant model, which makes images several times faster and uses far less VRAM than the full model, with quality that stays close. This guide sets up a fresh ComfyUI with the Nunchaku wheel, the ComfyUI-nunchaku node and the INT4 Flux models.
First published March 2025 on the old PixelAI Labs blog. Refreshed in October 2026: the steps below are checked, the links point to the current project pages, and newer options are mentioned where they help.
The project has moved to nunchux-ai/nunchaku and supports newer versions of Python, PyTorch and more models since this guide was written. Always pick the wheel that matches your own Python and PyTorch versions.
1. Check Python and PyTorch
python --version
nvcc --versionThe original setup used Python 3.11 with PyTorch 2.6 and CUDA 12.4.
2. Install ComfyUI
git clone https://github.com/Comfy-Org/ComfyUI.git
cd ComfyUI
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124
pip install -r requirements.txt3. Install the Nunchaku wheel
Pick the wheel for your Python and PyTorch version from the Nunchaku releases. The original guide used this one (PyTorch 2.6, Python 3.11, Windows):
pip install https://huggingface.co/mit-han-lab/nunchaku/resolve/main/nunchaku-0.1.4%2Btorch2.6-cp311-cp311-win_amd64.whl4. Install the custom nodes
cd custom_nodes
git clone https://github.com/Comfy-Org/ComfyUI-Manager comfyui-manager
git clone https://github.com/nunchux-ai/ComfyUI-nunchaku
cd ComfyUI-nunchaku
pip install -r requirements.txtBuilding an AI influencer? The UltraRealistic AI Influencer Pro course shows the complete system with today's models: design one character, train her own LoRA, and make KREA 2 photos and MiniMax H3 talking videos, on the cloud or your own GPU.
5. Download the models
Run these from the main ComfyUI folder:
pip install -U "huggingface_hub[cli]"
huggingface-cli download comfyanonymous/flux_text_encoders clip_l.safetensors --local-dir models/text_encoders
huggingface-cli download comfyanonymous/flux_text_encoders t5xxl_fp16.safetensors --local-dir models/text_encoders
huggingface-cli download black-forest-labs/FLUX.1-schnell ae.safetensors --local-dir models/vae
huggingface-cli download mit-han-lab/svdq-int4-flux.1-dev --local-dir models/diffusion_models/svdq-int4-flux.1-dev
huggingface-cli download mit-han-lab/svdq-int4-flux.1-fill-dev --local-dir models/diffusion_models/svdq-int4-flux.1-fill-dev6. Generate
Start ComfyUI, load an example workflow from the ComfyUI-nunchaku node (in its example_workflows folder) and pick the SVDQuant model in the Nunchaku loader. A first prompt to test the speed:
candid phone photo of a young woman laughing at a cafe table by the window, natural daylight, realistic skin textureFAQ
What is SVDQuant?
A 4-bit quantization method from MIT Han Lab that keeps quality close to the full model while cutting memory and making generation much faster.
Which Nunchaku wheel do I need?
The one that matches your Python version (cp311, cp312...) and PyTorch version. A wrong wheel fails to import.
Does it work with LoRAs?
Yes, the ComfyUI-nunchaku node pack includes a LoRA loader for its models. Check the node's examples for current support.
Take it further
Fast, low-VRAM Flux makes testing prompts cheap. If you want a ready-made, tested system instead of assembling one yourself, the UltraRealistic AI Influencer Pro course walks through character design, LoRA training, KREA 2 photos and MiniMax H3 videos step by step, and the InstantClone AI Influencer Suite creates consistent images and motion videos from one character image. More free guides are on the PixelAI Labs blog.
