This repository provides Docker build configurations that package ComfyUI together with all prerequisites, in a few different flavors. It also contains GitHub Actions definitions that can build and publish them to the GitHub Container Registry.
- Running on a GPU cloud provider like RunPod, QuickPod, Vast.ai or TensorDock.
- Running locally in a stable and isolated environment.
The images are ready-to-run with all necessary dependencies already installed. This means that upgrading to a new version of ComfyUI simply means downloading a new image, instead of updating existing files in place. So no fear of breaking an existing installation while updating!
The ComfyUI repository is checked periodically for updates, to ensure the images stay up to date.
| latest | master | cpu-latest | cpu-master | amd-latest | amd-master | |
|---|---|---|---|---|---|---|
base |  |  |  |  |  |  |
extensions |  |  |  |  |  |  |
ssh |  |  |  |  |  |  |
These images are currently published to the GitHub Container Registry:
| Image | Additional installed components |
|---|---|
comfyui-base | xFormers, FlashAttention-2, SageAttention2++, Nunchaku |
comfyui-extensions | ComfyUI-Manager |
comfyui-ssh | OpenSSH server |
| Tag | Description |
|---|---|
latest | Latest tagged release of ComfyUI for NVIDIA / CUDA 12.9 |
master | Latest commit of the ComfyUI master branch for NVIDIA / CUDA 12.9 |
amd-latest / amd-master | As above, but for AMD / ROCm 6.4 |
cpu-latest / cpu-master | As above, but a plain Ubuntu base image without GPU support |
docker run --gpus=all -p 8188:8188 ghcr.io/radiatingreverberations/comfyui-extensions:latest
Additional arguments will be forwarded to ComfyUI. For example, to enable SageAttention:
docker run --gpus=all -p 8188:8188 ghcr.io/radiatingreverberations/comfyui-extensions:latest --use-sage-attention
Without any additional configuration, any files created from ComfyUI will be lost whenever the container is removed or recreated, such as when updating to a new version. So when running locally, you most likely want certain directories to be located outside of the container. For persistent storage, mount these directories:
| Container Path | Purpose |
|---|---|
/comfyui/models | Model files |
/comfyui/user/default | User settings |
/comfyui/input | Input images/videos |
/comfyui/output | Generated outputs |
When using the extensions image that includes ComfyUI-Manager, you may also want the custom nodes you install to persist across Docker image updates. Mount this additional directory:
| Container Path | Purpose |
|---|---|
/comfyui/custom_nodes | Custom nodes |
Note that it is not recommended to update ComfyUI itself using ComfyUI-Manager - update to a later version of the Docker image instead.
An example of putting it all together using Docker Compose:
services:
comfyui:
image: ghcr.io/radiatingreverberations/comfyui-extensions:latest
container_name: comfyui
command:
- --use-sage-attention
ports:
- "8188:8188"
volumes:
- ./models:/comfyui/models
- ./user:/comfyui/user/default
- ./input:/comfyui/input
- ./output:/comfyui/output
- ./custom_nodes:/comfyui/custom_nodes
deploy:
resources:
reservations:
devices:
- driver: nvidia
count: all
capabilities: [gpu]
restart: unless-stopped
When running Docker images on Windows, it most likely runs on a Linux VM under WSL2. This means that they cannot access the Windows file systems directly, but have to use drvfs. This can make models take longer to load. The workaround is to put such files on a Linux file system directly inside WSL2.
If you were to simply open the ComfyUI port on your container, anyone on the internet will be able to connect, and all the traffic between your computer and the cloud provider would be unencrypted. One solution to this is to use the ssh image.
The ssh image starts an OpenSSH server on port 2222. So when running on a cloud provider, you would typically want to run a command like this:
docker run --gpus=all -p 2222:2222 ghcr.io/radiatingreverberations/comfyui-ssh:latest
The ssh image will display additional details on how to connect to the instance:
================================================================================ ComfyUI + SSH Tunnel ================================================================================ User: u-f0f1c7f3c8d148548dc4875c330849 SSH Port: 2222 Host key ID: SHA256:svSLAqHOM1w/Z2K9vSkssXHUcp6+XVtrqyUp4Wfgres Public IPv4: 123.456.789.101 How to connect: ssh -p 2222 u-f0f1c7f3c8d148548dc4875c330849@123.456.789.101 -L 8188:127.0.0.1:8188 Note! The actual IP address and port you need to connect to may be different depending on your hosting provider. Check their dashboard for the correct values if the above does not work. ================================================================================
By default the image will randomly generate a username and display it at startup. As this username is only known to you, no one else will be able to connect. It is also possible to configure the username by setting the SSH_USER environment variable:
docker run --gpus=all -e SSH_USER=u-f0f1mysecretuser0849 -p 2222:2222 ghcr.io/radiatingreverberations/comfyui-ssh:latest
This way you will not need to look at the console output to find it. To remain secure, ensure that the username you configure is not easy to guess. Alternatively, use key authentication by specifying your public key with SSH_KEY:
docker run --gpus=all -e SSH_USER=me -e SSH_KEY="ssh-ed25519 AAA...Qma" -p 2222:2222 ghcr.io/radiatingreverberations/comfyui-ssh:latest
Or even a password using SSH_PASSWORD:
docker run --gpus=all -e SSH_USER=me -e SSH_PASSWORD=extra-super-secret -p 2222:2222 ghcr.io/radiatingreverberations/comfyui-ssh:latest
Note: The published SSH image uses a static (pinned) host key. Its SHA256 fingerprint is printed at startup; verify it on first connect and expect it to remain unchanged across restarts and updates.
A simple script for downloading model files on the host is provided in the ssh image. After connecting over SSH, save the workflow you want to use from the UI. Then run something like:
download_workflow_models.py user/default/workflows/t2i-qwen.json
This works with workflows containing model links, such as those provided as templates in ComfyUI.
Instead of using the pre-built images it is also possible to build them locally.
docker buildx bake
NVIDIA CUDA runtime image: 12.9.1-runtime-ubuntu24.04, Python 3.12, git and additional components:
AMD ROCm runtime image: 6.4.1-dev-ubuntu24.04
