This is a Docker image for ComfyUI, which makes it extremely easy to run ComfyUI on Linux and Windows WSL2. The image also includes the ComfyUI Manager extension.
To get started, you have to install Docker. This can be either Docker Engine, which can be installed by following the Docker Engine Installation Manual or Docker Desktop, which can be installed by downloading the installer for your operating system.
To enable the usage of NVIDIA GPUs, the NVIDIA Container Toolkit must be installed. The installation process is detailed in the official documentation
The ComfyUI Docker image is available from the GitHub Container Registry.
Docker Compose is as simple as creating a docker-compose.yml file where you want to save your volumes (files).
services:
stable-diffusion-comfyui:
image: ghcr.io/therealpsv/comfyui-docker:latest
container_name: stable-diffusion-comfyui
volumes:
- ./sd-comfyui-data/models:/opt/comfyui/models
- ./sd-comfyui-data/custom_nodes:/opt/comfyui/custom_nodes
- ./sd-comfyui-data/pipcache:/root/.cache/pip
- ./sd-comfyui-data/input:/opt/comfyui/input
- ./sd-comfyui-data/userdata:/opt/comfyui/user/default
- ./sd-comfyui-data/customrequirements:/opt/customrequirements
- ./sd-comfyui-data/output:/opt/comfyui/output
stop_signal: SIGKILL
tty: true
restart: unless-stopped
ports:
- "8188:8188"
environment:
- PUID=1000
- PGID=1000
- CLI_ARGS=--disable-auto-launch
deploy:
resources:
reservations:
devices:
- driver: nvidia
count: all
capabilities:
- gpu
- compute
- utility
After saving the file you can simply run docker compose up -d to run the container.
| Folder | Description |
|---|---|
| models | Where your model files are stored. |
| custom_nodes | Where your custom nodes are stored. |
| output | Where your output files are stored. |
| input | Where any input files are saved by ComfyUI. |
| userdata | Where your settings for ComfyUI and ComfyUI Manager are stored, as well as your workflows. |
| customrequirements | ComfyUI Manager's "Install PIP Module" option doesn't work properly in this container. Instead, you can place your desired modules in a requirements.txt file in this folder, matching the pip requirements file format (typically just the name of the module, one per line), and the container will install it for you as it boots up. |
| pipcache | Caches saved Python PIP Modules to disk to speed up restarts. |
| Env Variable | Description |
|---|---|
| CLI_ARGS | Any arguments you want to pass to ComfyUI when it runs. |
| PUID | The user id of your host machine's linux user, if you want the container to attempt to use your user so the files in volumes are owned by your user. You can find this by running id -u. |
| PGID | The group id of your host machine's linux user, if you want the container to attempt to use your user so the files in volumes are owned by your user's group. You can find this by running id -g. |
Installing ComfyUI is as simple as pulling the image and starting a container, which can be achieved using the following command:
docker run \
--name comfyui \
--detach \
--restart unless-stopped \
--env PUID="$(id -u)" \
--env PGID="$(id -g)" \
--volume "<path/to/models/folder>:/opt/comfyui/models:rw" \
--volume "<path/to/custom/nodes/folder>:/opt/comfyui/custom_nodes:rw" \
--publish 8188:8188 \
--runtime nvidia \
--gpus all \
ghcr.io/therealpsv/comfyui-docker:latestPlease note, that the <path/to/models/folder> and <path/to/custom/nodes/folder> must be replaced with paths to directories on the host system where the models and custom nodes will be stored, e.g., $HOME/.comfyui/models and $HOME/.comfyui/custom-nodes, which can be created like so: mkdir -p $HOME/.comfyui/{models,custom-nodes}.
The --detach flag causes the container to run in the background and --restart unless-stopped configures the Docker Engine to automatically restart the container if it stopped itself, experienced an error, or the computer was shutdown, unless you explicitly stopped the container using docker stop. This means that ComfyUI will be automatically started in the background when you boot your computer. The two --env arguments inject the user ID and group ID of the current host user into the container. During startup, a user with the same user ID and group ID will be created, and ComfyUI will be run using this user. This ensures that files written to the volumes (e.g., models and custom nodes installed with the ComfyUI Manager) will be owned by the host system's user. Normally, the user inside the container is root, which means that the files that are written from the container to the host system are also owned by root. If you have run ComfyUI Docker without setting the environment variables, then you may have to change the owner of the files in the models and custom nodes directories: sudo chown -r "$(id -un):$(id -gn)" <path/to/models/folder> <path/to/custom/nodes/folder>. The --runtime nvidia and --gpus all arguments enable ComfyUI to access the GPUs of your host system. If you do not want to expose all GPUs, you can specify the desired GPU index or ID instead.
After the container has started, you can navigate to localhost:8188 to access ComfyUI.
If you want to stop ComfyUI, you can use the following commands:
docker stop comfyui
docker rm comfyuiWarning
While the custom nodes themselves are installed outside of the container, their requirements are installed inside of the container. This means that stopping and removing the container will remove the installed requirements. When the container is started again, the requirements will be automatically installed, but this may, depending on the number of custom nodes and their requirements, take some time. Using the pipcache volume speeds up the process by saving nodes to your local storage, so it doesn't have to redownload them.
In the folder with your docker-compose.yml file, simply run the following command:
docker compose pull && docker compose down && docker compose up -d
To update ComfyUI Docker to the latest version you have to first stop the running container, then pull the new version, optionally remove dangling images, and then restart the container:
docker stop comfyui
docker rm comfyui
docker pull ghcr.io/therealpsv/comfyui-docker:latest
docker image prune # Optionally remove dangling images
docker run \
--name comfyui \
--detach \
--restart unless-stopped \
--env USER_ID="$(id -u)" \
--env GROUP_ID="$(id -g)" \
--volume "<path/to/models/folder>:/opt/comfyui/models:rw" \
--volume "<path/to/custom/nodes/folder>:/opt/comfyui/custom_nodes:rw" \
--publish 8188:8188 \
--runtime nvidia \
--gpus all \
ghcr.io/therealpsv/comfyui-docker:latestIf you want to use the bleeding edge development version of the Docker image, you can also clone the repository and build the image yourself:
git clone https://github.com/therealpsv/comfyui-docker.git
docker build --tag therealpsv/comfyui-docker:latest comfyui-dockerNow, a container can be started like so:
docker run \
--name comfyui \
--detach \
--restart unless-stopped \
--env USER_ID="$(id -u)" \
--env GROUP_ID="$(id -g)" \
--volume "<path/to/models/folder>:/opt/comfyui/models:rw" \
--volume "<path/to/custom/nodes/folder>:/opt/comfyui/custom_nodes:rw" \
--publish 8188:8188 \
--runtime nvidia \
--gpus all \
therealpsv/comfyui-docker:latestYou can browse different variations and versions of this image here: https://github.com/therealpsv/comfyui-docker/pkgs/container/comfyui-docker/versions?filters%5Bversion_type%5D=tagged
Currently, each version is built nightly with the latest stable release of ComfyUI and ComfyUI Manager if either is updated. The images are currently based on 3 versions of the official PyTorch image:
- 2.5.1-cuda11.8-cudnn9
- 2.5.1-cuda12.1-cudnn9
- 2.5.1-cuda12.4-cudnn9 (default if you choose the
latesttag)
If you'd like to pull the latest image based on a specific version of PyTorch, you can pull one of these tags:
2.5.1-cuda11.8-cudnn9-latest2.5.1-cuda12.1-cudnn9-latest2.5.1-cuda12.4-cudnn9-latest(same as thelatesttag)
Similarly, the latest master build of ComfyUI at time of building is available with the nightly tag.
Versions based on specific builds of ComfyUI and ComfyUI Manager are available as well, though not all combinations of ComfyUI and ComfyUI Manager are, only the latest versions at the time of the nightly build. The naming scheme for specific versions is <PyTorch version>-cui-<ComfyUI version>-cuim-<ComfyUI Manager version>. You can browse these tags through the link above.