ComfyUI Docker Images

Easy to use container images of ComfyUI, supporting use in CPU only mode (slow) and with AMD, Intel Arc and NVIDIA GPUs.

This guide assumed you already have a Docker host (Docker Desktop or a server) with the appropriate hardware support in place, if not these links may help you get started:

• Getting Started | Docker
• Enable GPU support | Docker Docs

Getting Started

To run ComfyUI:

1. Download compose/docker-compose.{TAG}.yml to a suitable directory on your Docker host.
   • Replace {TAG} with the cpu, intel or nvidia as appropriate tag for your system.
   • If you'd prefer to build the image locally, use compose/docker-compose.{TAG}.build.yml instead, you'll need to update services.comfyui.build.args and services.comfyui.build.tags to the ComfyUI version you wish to build.
   • NOTE: The AMD dependencies are too large (~23GB) to build the image on GitHub Actions. You MUST build the image locally using compose/docker-compose.amd.build.yml to run on AMD hardware.
2. Rename your chosen file to docker-compose.yml.
3. Open a terminal in the chosen directory.
4. Start the service by running:

docker compose up -d

4. Monitor the container's startup by running:

docker compose logs -f

5. Once To see the GUI go to: http://0.0.0.0:8188 is logged by the container, ComfyUI should be accessible at http://localhost:8188/ from the Docker host.

Installing Models and Other Files

Once ComfyUI is running, you will need to install at least one model and will likely want to add other content too.

To copy files into the persisted content volume use the following command syntax:

docker compose cp ./v1-5-pruned-emaonly.safetensors comfyui:/opt/content/models/checkpoints/

The directory structure of /opt/content is:

/opt/
├─ content/
│  ├─ custom_nodes/
│  ├─ input/
|  │  ├─ 3d/
│  ├─ models/
|  │  ├─ checkpoints/
|  │  ├─ clip/
|  │  ├─ clip_vision/
|  │  ├─ configs/
|  │  ├─ controlnet/
|  │  ├─ diffusers/
|  │  ├─ diffusion_models/
|  │  ├─ embeddings/
|  │  ├─ gligen/
|  │  ├─ hypernetworks/
|  │  ├─ loras/
|  │  ├─ photomaker/
|  │  ├─ style_models/
|  │  ├─ text_encoders/
|  │  ├─ unet/
|  │  ├─ upscale_models/
|  │  ├─ vae/
|  │  ├─ vae_approx/
│  ├─ output/
│  ├─ temp/
│  ├─ user/

Enabling HTTPS

To run ComfyUI over HTTPS, you will need an X.509 certificate and the corresponding private key, these need to be copied into the container's tls volume, then the container needs to be restarted.

In the below example, a certificate called ComfyUI has been obtained using certbot, adjust the source paths as required for the certificate being installed.

docker compose cp --follow-link /etc/letsencrypt/live/ComfyUI/fullchain.pem comfyui:/etc/ssl/private/cert.pem
docker compose cp --follow-link /etc/letsencrypt/live/ComfyUI/privkey.pem comfyui:/etc/ssl/private/key.pem
docker compose up -d --force-recreate

Once complete, ComfyUI should be accessible at https://localhost:8188/.

Network Access

By default, ComfyUI is only available on the docker host, to allow network access edit docker-compose.yml - replacing - 127.0.0.1:8188:8188 with - '8188:8188' - and restart the container by running:

docker compose up -d --force-recreate

Once complete, ComfyUI should be accessible at http://{docker host IP / FQDN}:8188/ (https://{docker host IP / FQDN}:8188/ if HTTPS has been enabled as above).

NOTE: If a host-based firewall is present on the Docker host, or a network firewall is between the docker host and clients, rules will need to be added to allow access.

More Details

Image Tags

Architecture

Each supported hardware architecture has an associated set of images tagged as follows, there is no latest tag for this image.

The supported architecture tags are:

• cpu - runs ComfyUI on the CPU, rather than GPU - slow but widely compatible.
• intel - runs ComfyUI on a supported Intel Arc graphics card.
• nvidia - runs ComfyUI on a supported NVIDIA graphics card.

NOTE: As above the AMD image must be build locally due to the size of its dependencies.

ComfyUI Version

Each imaged version of ComfyUI, has associated tags for each architecture, taking the cpu architecture as an example, these include:

• cpu - The latest release of ComfyUI for the cpu architecture.
• 0-cpu - The latest 0.. release of ComfyUI for the cpu architecture.
• 0.3-cpu - The latest 0.3.* release of ComfyUI for the cpu architecture.
• 0.3.31-cpu - Version 0.3.31 of ComfyUI for the cpu architecture.

Environment Variables

ComfyUI can be customised using the following environment variables, these can be configured by creating a file named comfyui.env with one variable per line in the format VARIABLE_NAME=VALUE, only variables being changed from their Default need to be present in the file.

Variable	Purpose	Default
CORS_HEADER	Set the Cross-Origin Resource Sharing (CORS) header to restrict access from other domains. Can be set to either * for unrestricted or a URL to limit CORS requests to that domain and protocol.	*
CPU_ONLY	Run on the CPU only - this will be slow and should be used with the cpu tag.	false
GPU_ONLY	Run all operations on the GPU only.	false
MAX_UPLOAD_MB	The maximum upload size in megabytes - when running behind a reverse proxy ensure to set this there too.	100
SPLIT_CROSS_ATTENTION	Enable Split Cross Attention - can help with some memory and speed issues	false
VRAM	The VRAM profile to run with, can be auto, high, normal, low or no.	auto

Volumes

The image uses three volumes:

• content for user-content such as models, uploads, output files ,etc.
• tls for enabling HTTPS - place the certificate with chain in cert.pem and the unencrypted private key in key.pem
• tmp for temporary files - you may want to set this up as a tmpfs mount for performance optimisation.