madtomy/comfy-ui repository overview

⁠🧠 ComfyUI — GPU Docker Image for CUDA / RTX 5080

A clean, GPU-ready Docker image for ComfyUI⁠, built for AI image-generation workflows with PyTorch, CUDA, and cuDNN support.

This image is designed for NVIDIA GPU deployments, including newer RTX 50xx / Blackwell cards, and is suitable for Docker, Docker Compose, and Docker Swarm stacks.

⁠🚀 Features

• ✅ Official ComfyUI cloned directly from upstream
• ✅ GPU-ready with NVIDIA CUDA support
• ✅ Built on PyTorch 2.9.0 + CUDA 12.8 + cuDNN 9
• ✅ Suitable for RTX 5080 / Blackwell-class GPUs
• ✅ Exposes ComfyUI on port 7860
• ✅ Persistent volume support for models, input, output, user data, and custom nodes
• ✅ Optional ComfyUI update on container startup
• ✅ Optional custom-node requirements installation
• ✅ Rollback-friendly version tags

⁠🏷️ Available Tags

Recommended production usage: pin a version tag such as v0.20.1 instead of relying only on latest.

⁠🛠️ Base Image

pytorch/pytorch:2.9.0-cuda12.8-cudnn9-runtime

This provides:

• PyTorch 2.9.0
• CUDA 12.8 runtime
• cuDNN 9
• NVIDIA GPU acceleration support

⁠📦 Components

⁠🌐 Exposed Port

7860

Access ComfyUI at:

http://localhost:7860

⁠⚙️ Environment Variables

⁠🧩 Volume Layout

Recommended persistent mounts:

⁠🧪 Run with Docker

docker run --gpus all --rm -it \
  -p 7860:7860 \
  -e NVIDIA_VISIBLE_DEVICES=all \
  -e NVIDIA_DRIVER_CAPABILITIES=compute,utility \
  -e PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \
  -v /path/to/models:/opt/ComfyUI/models \
  -v /path/to/input:/opt/ComfyUI/input \
  -v /path/to/output:/opt/ComfyUI/output \
  -v /path/to/user:/opt/ComfyUI/user \
  -v /path/to/custom_nodes:/opt/ComfyUI/custom_nodes \
  madtomy/comfy-ui:v0.20.1

⁠🧪 Run Latest Tag

docker run --gpus all --rm -it \
  -p 7860:7860 \
  -v /path/to/models:/opt/ComfyUI/models \
  -v /path/to/input:/opt/ComfyUI/input \
  -v /path/to/output:/opt/ComfyUI/output \
  -v /path/to/custom_nodes:/opt/ComfyUI/custom_nodes \
  madtomy/comfy-ui:latest

⁠🧰 Custom CLI Arguments

Example: enable extra ComfyUI arguments:

docker run --gpus all --rm -it \
  -p 7860:7860 \
  -e CLI_ARGS="--preview-method auto" \
  madtomy/comfy-ui:v0.20.1

⁠🔄 Startup Updates

By default, the container can update ComfyUI on startup:

-e COMFYUI_UPDATE_ON_START=true

For stable pinned deployments, disable runtime updates:

-e COMFYUI_UPDATE_ON_START=false

Recommended production pattern:

• Use pinned image tags for rollback safety, e.g. madtomy/comfy-ui:v0.20.1
• Test new tags before promoting them to latest
• Keep the previous known-good tag available for quick rollback

⁠🧱 Docker Compose Example

services:
  comfyui:
    image: madtomy/comfy-ui:v0.20.1
    ports:
      - "7860:7860"
    environment:
      NVIDIA_VISIBLE_DEVICES: all
      NVIDIA_DRIVER_CAPABILITIES: compute,utility
      PYTORCH_CUDA_ALLOC_CONF: expandable_segments:True
      COMFYUI_UPDATE_ON_START: "false"
      COMFYUI_INSTALL_CUSTOM_NODE_REQUIREMENTS: "false"
      CLI_ARGS: ""
    volumes:
      - /path/to/models:/opt/ComfyUI/models
      - /path/to/input:/opt/ComfyUI/input
      - /path/to/output:/opt/ComfyUI/output
      - /path/to/user:/opt/ComfyUI/user
      - /path/to/custom_nodes:/opt/ComfyUI/custom_nodes
      - /path/to/cache:/root/.cache
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]

⁠🐳 Docker Swarm Notes

For Swarm deployments, GPU scheduling depends on your node configuration. This image is GPU-ready, but the Swarm node must have:

• NVIDIA driver installed
• NVIDIA Container Toolkit configured
• GPU resource reservation or node placement configured correctly

Example service image reference:

image: madtomy/comfy-ui:v0.20.1

⁠🔙 Rollback

If a new image or ComfyUI update causes issues, redeploy the previous known-good tag:

docker pull madtomy/comfy-ui:v0.20.1

Then update your Compose/Swarm stack back to that tag.

For safer production behavior, set:

COMFYUI_UPDATE_ON_START=false

⁠🧾 License

This image includes ComfyUI⁠, which is licensed under the MIT License.

Please refer to the upstream ComfyUI repository for full license details.

⁠💡 Maintainer

Maintained by: madtomy⁠
Source: ComfyUI on GitHub⁠