ComfyUI/.github at master · oxc/ComfyUI · GitHub
Skip to content

.github

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History
 
 

.github

Folders and files

README.md

ComfyUI Docker Images

Automatically built, always up-to-date Docker images of ComfyUI.

This repository is a lightweight fork of ComfyUI whose only purpose is to build and publish Docker images. It tracks the upstream master branch (rebased several times a day) and adds nothing but a Dockerfile and the CI needed to publish images for the common GPU backends.

Not affiliated with the ComfyUI project. For everything about ComfyUI itself — features, models, workflows, docs and support — please head to the upstream repository and comfy.org. Upstream has decided not to maintain an official Dockerfile for now, which is why these images live here instead.

What is ComfyUI?

ComfyUI is a powerful and modular node-graph engine for generative AI. It lets you build image, video, audio and 3D generation pipelines by wiring together nodes, with native support for the latest open-source models (SD/SDXL/SD3, Flux, Qwen Image, Wan, Hunyuan, and many more). See the upstream README for the full picture.

Images

Images are published to the GitHub Container Registry:

ghcr.io/oxc/comfyui

A separate image variant is built for each supported PyTorch backend. Pick the one that matches your hardware:

Variant tag suffix Backend Use for
-cu130 CUDA 13.0 NVIDIA GPUs (latest)
-cu128 CUDA 12.8 NVIDIA GPUs
-cu126 CUDA 12.6 NVIDIA GPUs (older drivers)
-rocm7.2 ROCm 7.2 AMD GPUs
-cpu CPU only No GPU (slow)

Tags

Every variant is published under several tags:

  • latest-<variant> — the newest published release, e.g. latest-cu130
  • <version>-<variant> / <major>.<minor>-<variant> — a specific ComfyUI release
  • master-<variant> — the tip of upstream master (bleeding edge)
  • sha-<commit>-<variant> — a specific commit
# NVIDIA
docker pull ghcr.io/oxc/comfyui:latest-cu130
# AMD
docker pull ghcr.io/oxc/comfyui:latest-rocm7.2
# CPU only
docker pull ghcr.io/oxc/comfyui:latest-cpu

Running

ComfyUI listens on port 8188. Mount volumes for your models and I/O so they survive container restarts.

NVIDIA

Requires the NVIDIA Container Toolkit.

docker run --rm -it \
  --gpus all \
  -p 8188:8188 \
  -v "$PWD/models:/app/models" \
  -v "$PWD/input:/app/input" \
  -v "$PWD/output:/app/output" \
  -v "$PWD/user:/app/user" \
  ghcr.io/oxc/comfyui:latest-cu130

AMD (ROCm)

docker run --rm -it \
  --device /dev/kfd --device /dev/dri \
  --group-add video \
  --security-opt seccomp=unconfined \
  -p 8188:8188 \
  -v "$PWD/models:/app/models" \
  -v "$PWD/input:/app/input" \
  -v "$PWD/output:/app/output" \
  -v "$PWD/user:/app/user" \
  ghcr.io/oxc/comfyui:latest-rocm7.2

CPU only

docker run --rm -it \
  -p 8188:8188 \
  -v "$PWD/models:/app/models" \
  -v "$PWD/input:/app/input" \
  -v "$PWD/output:/app/output" \
  -v "$PWD/user:/app/user" \
  ghcr.io/oxc/comfyui:latest-cpu

Then open http://localhost:8188.

docker compose

services:
  comfyui:
    image: ghcr.io/oxc/comfyui:latest-cu130
    ports:
      - "8188:8188"
    volumes:
      - ./models:/app/models
      - ./input:/app/input
      - ./output:/app/output
      - ./user:/app/user
      # persist packages installed by custom nodes (see below)
      - ./custom_nodes:/app/custom_nodes
      - custom_venv:/app/custom_venv
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]

volumes:
  custom_venv:

Configuration

The container is configured through environment variables:

Variable Default Description
COMFYUI_ADDRESS 0.0.0.0 Address ComfyUI binds to
COMFYUI_PORT 8188 Port ComfyUI listens on
COMFYUI_EXTRA_ARGS (empty) Extra arguments appended to main.py, e.g. --preview-method taesd

For example, to enable high-quality previews:

docker run ... -e COMFYUI_EXTRA_ARGS="--preview-method taesd" ghcr.io/oxc/comfyui:latest-cu130

Custom nodes

Custom nodes often install their own Python dependencies. To keep those installs across container restarts, mount a volume at /app/custom_venv. On startup the container copies the baked-in virtual environment into that volume (if empty) and uses it from then on, so any packages installed by ComfyUI-Manager or by hand persist. Mount /app/custom_nodes alongside it to keep the nodes themselves.

Building yourself

The images are built from the Dockerfile in this repo. Pass a PYTORCH_INSTALL_ARGS build arg to select the backend:

# NVIDIA
docker build --build-arg PYTORCH_INSTALL_ARGS="--index-url https://download.pytorch.org/whl/cu130" -t comfyui .

# AMD
docker build --build-arg PYTORCH_INSTALL_ARGS="--index-url https://download.pytorch.org/whl/rocm7.2" -t comfyui .

# CPU only (also bake in the --cpu flag)
docker build \
  --build-arg PYTORCH_INSTALL_ARGS="--index-url https://download.pytorch.org/whl/cpu" \
  --build-arg EXTRA_ARGS=--cpu \
  -t comfyui .

BuildKit is required (it is the default on modern Docker; otherwise set DOCKER_BUILDKIT=1).

How the images stay current

A scheduled GitHub Action rebases this fork onto upstream master several times a day. Each successful sync — and every upstream release — triggers the image build workflow, so the published tags follow upstream closely without any manual work.

License

ComfyUI is licensed under the GPL-3.0 License; this fork inherits it. All credit for ComfyUI goes to the ComfyUI authors and contributors.