For those wondering if they need an Nvidia GPU for local image generation, the answer is no.
Yes, not having CUDA is a big disadvantage but the community and Intel have worked hard to make up for this difference. Enough so that you can leverage the full power of your GPU (Assuming you know how to balance the workload).
An ARC A770 16GB with 32GB of system RAM powers my setup, I understand most people have much less VRAM but VRAM/RAM is the main limiting factor here especially for speed but also for your ability to generate at all.
I hope this guide can assist any ARC users out there, it's not perfect, but I want to share what I know to help those running into the same issues.
Latest Installation for Windows 10 as of (26-12-2025):
Install the latest Intel ARC drivers and Intel Graphics software:
https://www.intel.com/content/www/us/en/products/docs/discrete-gpus/arc/software/drivers.html
ComfyUI win-portable XPU:
This is a much more user friendly option than the previous v1 instructions. I recommend this for beginners.
https://github.com/YanWenKun/ComfyUI-WinPortable-XPU?tab=readme-ov-file
Python 3.10-3.12 (I'm using 3.12):
https://www.python.org/downloads/release/python-31210/
When installing python, make sure to enable python as a PATH environment variable
(For help: https://www.geeksforgeeks.org/python/how-to-add-python-to-windows-path/)
Run the "Run_Launcher.bat", it will auto install ComfyUI Manager and automatically open the front end.
Below is the v1 version I had posted previously.
Installation for Windows 10 as of (20-11-2025):
ComfyUI for Intel ARC:
Python 3.10-3.12 (I'm using 3.12):
https://www.python.org/downloads/release/python-31210/
When installing python, make sure to enable python as a PATH environment variable
(For help: https://www.geeksforgeeks.org/python/how-to-add-python-to-windows-path/)
Intel extension for Pytorch:
(DirectML is also an alternative, not listed here)
Note for the future. Intel Extension for Pytorch will reach EOL by the end of March 2026. Use PyTorch directly after this.
Intel oneapi base toolkit:
https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit-download.html
Install the latest Intel ARC drivers and Intel Graphics software:
https://www.intel.com/content/www/us/en/products/docs/discrete-gpus/arc/software/drivers.html
Git bash: (I use it to launch ComfyUI but you can also use CMD prompt or something else)
install pip from inside git bash, just launch it and type: "sudo easy_install pip" (no quotes, press enter, then update pip using "python - m pip install --upgrade pip)
-When opening your terminal of choice, you need to navigate to your ComfyUI folder usually found in C:\Users\[USERNAMEHERE]\ComfyUI, in git bash entering "cd ComfyUI" (no quotes needed for these commands) will navigate you to your ComfyUI folder.
If this is your first time setting up ComfyUI you need to install the requirements, you already have the file, you just need to tell git bash to get pip to install it, this is a one time process "pip install -r requirements.txt"
(OPTIONAL BUT RECOMMENDED: The requirements.txt file in the ComfyUI you downloaded earlier is quite out of date, I recommend downloading and placing this "requirements-ipex.txt" file at the bottom of this discussion page into your ComfyUI folder https://github.com/comfyanonymous/ComfyUI/discussions/8626)
(replace the "requirements.txt" section with requirements-ipex.txt so instead of "pip install -r requirements.txt" type "pip install -r requirements-ipex.txt" ) -requirements-ipex.txt should be modified in notepad++ to manually update the file, telling it to grab the latest versions.
[End of optional section]
-Once installed you can enter "python main.py" and after pressing enter ComfyUI will start up, it can take a bit, once successfully started up you will see an ip address close to the bottom of the terminal, enter this ip into your browser to navigate into the UI.
ComfyUI Manager (for simplicity of use for installing missing nodes):
https://github.com/Comfy-Org/ComfyUI-Manager
I also recommend installing a system monitoring tool (hwinfo, GPUZ, etc), you need to understand how your system is handling the loads you're throwing at it. Just because it is running, does not mean it is running well.
-END OF V1-
Tips for speed and usage:
-Experiment with different workflows, find the one that's right for you!
-When loading a model, it will be loaded into RAM before passed into VRAM, if it all fits in VRAM, your generations will be fast and power hungry. If not and there is spill over into RAM, it will dramatically slow down your generations. (If something is showing 8GB of VRAM used, it may be 7.1GB in VRAM and 0.9GB in RAM acting as graphics memory, you will still feel this slow down)
-More loras? More VRAM used, just load what you need
-Find a model/ckpt/gguf etc. That fits into your VRAM with some space to spare. You probably see multiple versions of models that have 4Q, 5Q, 6Q, etc. This is the quantized bit version of the original model, the lower the bit, the less VRAM it takes up.
-Minimize open programs, set your browser to not use GPU acceleration to minimize VRAM usage. I am at around 488MB of VRAM usage before I start a generation.
-In intel graphics software, go to performance and click tuning, here you can set the power limit much higher (228W for me), this will generate a lot of heat.
-CLIPS can be loaded to your CPU in the workflow (If you find yourself short on VRAM)
-The higher the step count, the longer it will take to generate, but the quality is usually higher to a point.
-Yes, WAN and Hunyuan video generation works on Intel ARC GPUs.
-If you're wondering why AI video generations are so short, VRAM VRAM VRAM, everything revolves around your VRAM and RAM.
-All models, checkpoints, loras, etc will be place into your ComfyUI folder in its respective spot, (ie: loras go into \\ComfyUI\models\loras), you will need to make sure they are place in the correct folder and point ComfyUI workflows to them.
AI generations on tiny computers like ours is a balancing act, you need to configure, monitor and maintain your system for what you want to achieve balanced off of what your system can handle.
Happy Generating!