ComfyUI_Docker

Minimal, stateless Docker environment for ComfyUI.

Built on python:3.10-slim. Bypasses bloated NVIDIA base images by relying on PyTorch's bundled CUDA runtime. Designed for persistence across container destruction, utilizing local volumes for models, plugins, and pip cache.

Structure

• Stateless base: The image contains only the OS, Python, ComfyUI, and ComfyUI-Manager.
• Persistent cache: Mounting ~/.cache/pip prevents redownloading dependencies for custom nodes when containers are recreated.
• Dynamic entrypoint: An initialization script symlinks ComfyUI-Manager and resolves requirements.txt for custom nodes automatically on startup.

Usage

Create docker-compose.yaml:

services:
  comfyui:
    image: git.trashlab.qzz.io/unins0/comfyui_docker:latest
    ports:
      - "8188:8188"
    volumes:
      - ./models:/app/models
      - ./custom_nodes:/app/custom_nodes
      - ./user:/app/user
      - ./output:/app/output
      - ./pip_cache:/root/.cache/pip
    # Use ["--cpu"] for CPU-only
    # command: ["--cpu"]
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]

Commands

docker compose up -d   # Start daemon
docker compose stop    # Suspend
docker compose start   # Resume
docker compose down    # Destroy container (data persists in volumes)

Credits

This project integrates methods from the following repositories:

• YanWenKun/ComfyUI-Docker - Slim image methodology and manager cache bypass.
• lecode-official/comfyui-docker - Symlink entrypoint implementation for ComfyUI-Manager persistence.
• Google Gemini

License

GPLv3. See LICENSE.