ComfyUI Intel XPU

Docker image for running ComfyUI with native PyTorch XPU acceleration on Intel Arc GPUs.

This image was created primarily to make running ComfyUI on Intel Arc GPUs — including Intel Arc Pro Battlemage GPUs — straightforward on Docker and Unraid without requiring GPU passthrough to a virtual machine.

Features

• Native PyTorch XPU acceleration
• Intel Level Zero GPU runtime
• Intel OpenCL runtime
• ComfyUI
• ComfyUI Manager
• Hugging Face CLI (hf)
• aria2 for fast/resumable downloads
• Persistent model storage
• Persistent custom nodes
• Persistent input/output directories
• Automatic creation of standard ComfyUI model directories
• Designed for Intel GPUs using the Linux xe driver
• Tested on Intel Arc Pro B70
• Optional ComfyUI startup arguments through CLI_ARGS
• Optional runtime Python packages through PIP_PACKAGES

Tested Configuration

The initial release has been validated with:

The B70 was detected by PyTorch as:

XPU available: True
XPU device count: 1
XPU 0: Intel(R) Graphics [0xe223]

ComfyUI reported approximately 31 GB of usable VRAM:

Total VRAM 31023 MB
Device: xpu:0 Intel(R) Graphics [0xe223]

Docker Image

heroeswearkapes/comfyui-intel-xpu:latest

Versioned images are also published using the build date and upstream ComfyUI Git commit:

heroeswearkapes/comfyui-intel-xpu:YYYY.MM.DD-<commit>

Example:

heroeswearkapes/comfyui-intel-xpu:2026.08.13-b323a34

Host Requirements

The Intel GPU must already be functioning on the Docker host.

The container does not provide the Linux kernel GPU driver.

The host should provide:

• Intel Arc-compatible GPU
• Linux xe driver
• /dev/dri render device
• Docker

Verify the GPU driver with:

lspci -nnk | grep -A4 -Ei 'Intel|VGA|Display'

A working Intel Arc GPU should show something similar to:

Kernel driver in use: xe
Kernel modules: xe

Finding Your Intel GPU Render Device

Intel GPUs appear under:

/dev/dri/

List available render devices:

ls -la /dev/dri/

On systems with multiple GPUs, determine which render node belongs to which PCI device:

for r in /dev/dri/renderD*; do
    echo "=== $r ==="
    udevadm info -q property -n "$r" | grep -E "PCI_SLOT_NAME|DEVPATH"
    echo
done

Then compare the PCI address with:

lspci -nnk

For example, on the test system the Arc Pro B70 was:

87:00.0 Intel Corporation Battlemage G31 [Arc Pro B70]

and mapped to:

/dev/dri/renderD131

Do not assume your GPU will use renderD131.

Render device numbering varies between systems and may change when hardware configuration changes.

Unraid Installation

1. Create Storage

It is recommended to create an Unraid share for AI models and generated content.

The included Unraid template defaults to a share named:

AI

which produces paths such as:

/mnt/user/AI/comfyui/models
/mnt/user/AI/comfyui/input
/mnt/user/AI/comfyui/output

You may use any share you prefer. Simply change the Host Path values during container installation.

Application configuration and custom nodes default to:

/mnt/user/appdata/comfyui-intel-xpu/

2. GPU Device

Set the Intel GPU device to the appropriate render node for your system.

Example:

/dev/dri/renderD131

3. Web Interface

The default ComfyUI port is:

8188

Once running, access:

http://UNRAID-IP:8188

The Unraid WebUI button should also open ComfyUI automatically.

Docker CLI

Example:

docker run -d \
  --name comfyui-intel-xpu \
  --device=/dev/dri/renderD128:/dev/dri/renderD128 \
  --ipc=host \
  -p 8188:8188 \
  -v /path/to/config:/config \
  -v /path/to/custom_nodes:/custom_nodes \
  -v /path/to/models:/models \
  -v /path/to/input:/input \
  -v /path/to/output:/output \
  --restart unless-stopped \
  heroeswearkapes/comfyui-intel-xpu:latest

Replace:

/dev/dri/renderD128

with the render device belonging to your Intel GPU.

Docker Compose

Example:

services:
  comfyui:
    image: heroeswearkapes/comfyui-intel-xpu:latest
    container_name: comfyui-intel-xpu

    devices:
      - /dev/dri/renderD128:/dev/dri/renderD128

    ports:
      - "8188:8188"

    volumes:
      - ./data/config:/config
      - ./data/models:/models
      - ./data/input:/input
      - ./data/output:/output
      - ./data/custom_nodes:/custom_nodes

    environment:
      # Optional additional ComfyUI startup arguments
      # CLI_ARGS: "--disable-dynamic-vram --lowvram --cpu-vae --reserve-vram=1 --disable-smart-memory"

      # Optional additional Python packages
      # PIP_PACKAGES: "opencv-python imageio_ffmpeg"

    ipc: host

    restart: unless-stopped

Optional Runtime Configuration

Additional ComfyUI CLI Arguments

Additional ComfyUI command-line arguments can be supplied with the CLI_ARGS environment variable.

Example:

CLI_ARGS=--disable-dynamic-vram --lowvram --cpu-vae --reserve-vram=1 --disable-smart-memory

The supplied arguments are appended to the standard ComfyUI launch command.

On Unraid, this option is available under Advanced View as Additional ComfyUI CLI Arguments.

Common options include:

For the complete list of supported ComfyUI command-line arguments, see the ComfyUI CLI Argument Reference.

The available arguments depend on the version of ComfyUI included in the container. You can always view the exact options supported by your installed image with:

docker exec ComfyUI-Intel-XPU python /opt/ComfyUI/main.py --help

Additional Python Packages

Optional Python packages can be installed automatically at container startup using the PIP_PACKAGES environment variable.

Example:

PIP_PACKAGES=opencv-python imageio_ffmpeg

Multiple packages may be separated by spaces. Standard pip package specifications are supported.

Packages are installed into the container's Python environment before ComfyUI Manager and ComfyUI are started.

On Unraid, this option is available under Advanced View as Additional Python Packages.

Packages installed through PIP_PACKAGES are part of the running container. If the container is recreated or updated, the requested packages will be installed again automatically.

Model Directories

The container automatically creates common ComfyUI model directories on startup.

These include:

/models/
├── checkpoints/
├── clip/
├── clip_vision/
├── configs/
├── controlnet/
├── diffusers/
├── diffusion_models/
├── embeddings/
├── gligen/
├── hypernetworks/
├── loras/
├── model_patches/
├── photomaker/
├── style_models/
├── text_encoders/
├── unet/
├── upscale_models/
├── vae/
└── vae_approx/

For example, normal checkpoint models can be placed in:

/models/checkpoints/

On an Unraid installation using the recommended paths, that corresponds to:

/mnt/user/AI/comfyui/models/checkpoints/

Hugging Face

The official Hugging Face hf CLI is included.

Verify it with:

docker exec comfyui-intel-xpu hf --help

Models can be downloaded directly into persistent ComfyUI storage.

Example:

docker exec comfyui-intel-xpu \
  hf download stabilityai/stable-diffusion-xl-base-1.0 \
  sd_xl_base_1.0.safetensors \
  --local-dir /models/checkpoints

For gated or private models, provide a Hugging Face token using the HF_TOKEN environment variable or authenticate using the Hugging Face CLI.

Never bake your Hugging Face token into the Docker image.

aria2

aria2c is included for fast, resumable downloads.

Verify:

docker exec comfyui-intel-xpu aria2c --version

Example:

docker exec comfyui-intel-xpu \
  aria2c \
  -c \
  -x 8 \
  -s 8 \
  -d /models/checkpoints \
  "MODEL_DOWNLOAD_URL"

ComfyUI Manager

ComfyUI Manager is automatically installed into the persistent custom nodes directory:

/custom_nodes/comfyui-manager

Because /custom_nodes is persistent, Manager and other custom nodes survive container upgrades and recreation.

Manager can be used to install and manage additional ComfyUI custom nodes.

Updating ComfyUI

ComfyUI itself is intentionally managed by the Docker image.

You may see a message in Manager similar to:

Your ComfyUI isn't git repo.

This is expected.

Do not use Manager to update the core ComfyUI installation inside this container.

Instead, update the Docker image:

docker pull heroeswearkapes/comfyui-intel-xpu:latest

and recreate/restart the container.

This keeps the application image reproducible and prevents container-local ComfyUI modifications from being lost during upgrades.

Custom Node Compatibility

Not every ComfyUI custom node supports Intel XPU.

Custom nodes that explicitly require technologies such as:

• CUDA
• NVIDIA-specific libraries
• CUDA-only Triton kernels
• NVIDIA-specific Flash Attention implementations

may not function on Intel GPUs.

The core ComfyUI installation and standard PyTorch operations use native Intel XPU acceleration.

Troubleshooting

XPU is unavailable

Check the container log for:

XPU available: True

If it reports:

XPU available: False

first verify that the GPU device was passed into the container.

Example:

docker exec comfyui-intel-xpu ls -la /dev/dri

Then test PyTorch directly:

docker exec comfyui-intel-xpu python -c \
'import torch; print(torch.__version__); print(torch.xpu.is_available()); print(torch.xpu.device_count())'

Permission Problems

Verify the render device exists on the host:

ls -l /dev/dri/renderD*

The Docker container must have access to the selected render device.

View Logs

docker logs -f comfyui-intel-xpu

Successful Intel XPU initialization should resemble:

XPU available: True
XPU device count: 1
Total VRAM 31023 MB
Device: xpu:0 Intel(R) Graphics

Updating

Pull the latest image:

docker pull heroeswearkapes/comfyui-intel-xpu:latest

Then recreate the container using the same persistent volume mappings.

Unraid users can update through the normal Docker update mechanism.

Versioning

The latest tag points to the current validated build.

Versioned releases use:

YYYY.MM.DD-<ComfyUI commit>

This makes it possible to roll back to a known ComfyUI revision if an upstream change causes problems.

Disclaimer

This is an independent community Docker image.

It is not an official ComfyUI, Intel, PyTorch, Hugging Face, or Unraid project.

Intel GPU and custom-node compatibility can vary by GPU generation, kernel, driver, PyTorch version, and individual workflow.

Acknowledgements

Special thanks to MDKAOD for early community feedback and contributions to ComfyUI Intel XPU.

Their feature requests and pull request helped drive the addition of:

• Custom ComfyUI startup arguments through CLI_ARGS
• Optional runtime Python package installation through PIP_PACKAGES
• Improved runtime configurability for Docker and Unraid users

Thank you for taking the time to test the project, provide feedback, and contribute ideas and code back to the community.

License

This project is licensed under the MIT License.

See the LICENSE file for the full license text.