How to install "Nunchaku", a Plugin for ComfyUI?
The last week was rich in anouncements:
Wan 2.5, an AI tool to create videos can be tested online.
Lucy-Edit-Dev, an AI tool to edit videos, has been released
(the 14b is just online and costs credits, while the 5b is usable offline).
Wan2.2-Animate, an AI tool to create new videos from an old one using a prompt and a picture, has been released.
and finally Qwen-Image-Edit-2509, a new version of the tool to edit pictures using a prompt has been released.
Good News ! These tools are all free, at least partially...
If you try to use some AI models or some workflows in ComfyUI, you may have to install "Nunchaku", a Plugin for ComfyUI.
For example, if you want to test Qwen-Image-Edit-2509 and you don't have so much VRAM, you may tumble upon the Nunchaku version of Qwen-Image-Edit-2509.
This version has the "lightning" lora v2 baked-in, to reduce the number of steps to edit the pictures.
Also the workflow it uses for "Qwen image Edit 2509" has a feature (the "num_block_on_gpu" setting in the model loader) that makes it a good candidate, if you don't have so much video ram (vram): you know the stuff that you need to load the AI models...
Additionally and that's my own opinion, it seems to respect more the original picture than for example the gguf version of "Qwen image Edit 2509" (if you are new to this world, GGUF is some format to reduce the size of the models).
This doesn't mean that Qwen-Image-Edit-2509, Nunchaku version, can suddenly do everything and your prompts will likely make the difference.
For example, I am still at trying to find a way to replace an object, by a similar object in another picture...
The installation process of the Nunchaku nodes is a little bit unusual and when i tried to install it using the usual nodes installer, i received a "NunchakuQwenImageDiTLoader missing" error, when I try to use one of the example workflow of "Nunchaku".
To install this node, you have to read carefully what is written here:
https://nunchaku.tech/docs/ComfyUI-nunchaku/get_started/installation.html#step-2-install-the-nunchaku-backend
(It is easy, right ? Just read the manual ?...)
- Make sure that ComfyUi has been updated to a recent version and make sure it has all the requirements to work properly.
- Use a command line interface (run >> cmd) or the powershell environment (windows)
- Make sure that the environment of ComfyUI is activated in your command line interface.
- go the "custom_nodes" of ComfyUI: ComfyUI\custom_nodes
- Make sure to use THIS command: git clone https://github.com/mit-han-lab/ComfyUI-nunchaku nunchaku_nodes
Note that the "nunchaku_nodes" part changes the name of the directory and is unusual, so pay attention.
- Then go the newly created directory:
cd nunchaku_nodes
and do this:
pip install -r requirements.txt
Do a: pip list
to check that your version of torch is still cuda based (example: 2.8.0+cu128)
If not do this:
pip install --force-reinstall torch torchvision --extra-index-url https://download.pytorch.org/whl/cu128
(in function of your version of CUDA, here: 12.8)
Finally, download and open this workflow:
https://github.com/nunchaku-tech/ComfyUI-nunchaku/blob/main/example_workflows/install_wheel.json
Since installing this version: Nunchaku version: 1.0.1.dev20250926, ComfyUI-nunchaku version: 1.0.1
Gave me later this answer in my console log:
ComfyUI-nunchaku 1.0.1 is not compatible with nunchaku 1.0.1.dev20250926. Please update nunchaku to a supported version in ['v1.0.0'].
I suggest to select the version: latest
And the dev_version: none
Mode: install
Click on "run" like to start a workflow.
This will install the correct "wheel" file (whl) for ComfyUI.
If it returns an error, check the location of the whl files in the console log:
It it something like that: https://github.com/nunchaku-tech/nunchaku/releases/download/v1.0.0/nunchaku-1.0.0+torch2.8-cp310-cp310-win_amd64.whl
And follow what is written:
"place it in `custom_nodes/ComfyUI-nunchaku"
Then run again.
Restart ComfyUI and reload the browser.
If the package has already be installed once, it will persist to return an error.
If you have any problem, you could still delete the "nunchaku" libraries located here:
venv\Lib\site-packages
Finally, to use the Qwen-Image-Edit-2509 workflow, use this ready-made workflow:
https://github.com/nunchaku-tech/ComfyUI-nunchaku/blob/main/example_workflows/nunchaku-qwen-image-edit-2509-lightning.json
Note that I didn't see any difference of nodes between this workflow and the other one that is without "lightning".