rncz.net

ComfyUI on Windows 11 with ROCm: Flux.1-dev on AMD GPUs

comfyuirocmamdpytorchfluximage-generationwindows

NVIDIA has the easy path, but AMD GPUs can run stable diffusion too — with the official ROCm nightly release and PyTorch, you can run ComfyUI on Windows 11 and generate images with Flux.1-dev. This works on the RX 7000, RX 8000, and RX 9000 series as well as Strix Halo.

This post walks through the full setup: drivers, Git, Python, ComfyUI, the ROCm nightly PyTorch wheels, and finally generating images with Flux.1-dev.

1. Install the AMD drivers

Make sure you have the right drivers for your AMD GPU. Go to Google, search for Adrenalin AMD, and download the Windows drivers from the link. Run the installer — it downloads everything it needs, and afterwards you can check in the Adrenalin software that your GPU is detected.

2. Install Git SCM

Git for Windows makes it easy to clone repositories. Download it from git-scm.com and run the installer. The defaults are fine for most things:

  • Editor: pick whatever you like (e.g. Notepad++).
  • Use Git from the command line.
  • Use the bundled OpenSSH.
  • Use the native Windows Secure Channel library for TLS.
  • Checkout: Checkout as-is.
  • Credentials: Git Credential Manager.

Then verify in a terminal (Win key, type cmd, Enter):

git --version

If a version is printed, you’re good.

3. Install Python 3.11 and 3.13

Go to python.org/downloads and grab the Windows installers for Python 3.11 (e.g. 3.11.9) and Python 3.13. For each one, tick “Add Python to the PATH” during installation and install them one after the other.

The installers add the py launcher, so you can target a specific version. I tested Python 3.11 and it works fine with ComfyUI, but Python 3.13 should work too — and it’s probably the better choice to use nowadays:

py -3.11 --version
py -3.13 --version

4. Clone ComfyUI

Create a working directory and clone the ComfyUI repository:

mkdir AI
cd AI
git clone https://github.com/comfyanonymous/ComfyUI
cd ComfyUI

5. Create a virtual environment

Use Python 3.13 to create a virtual environment for ComfyUI, then activate it:

py -3.13 -m venv venv
venv\Scripts\activate

Check that the right interpreter is active:

python --version

It should print Python 3.13.x. You can also use pip list to see which packages are installed in the environment.

6. Install the ROCm nightly + PyTorch

Head to the official ROCm GitHub repository and open the RELEASES.md file. First, check that your GPU is in the supported list (RX 7000 / 8000 / 9000, Strix Halo).

The page also lists a device-gfx... ID for every supported GPU — for example, the RX 7800 XT is gfx1100. Find the one that matches your GPU and use it in the pip command below (this example uses gfx942). Run it in your (activated) virtual environment:

pip install --index-url https://rocm.nightlies.amd.com/whl-multi-arch/ \
    "torch[device-gfx942]" "torchvision[device-gfx942]" torchaudio

This downloads all the PyTorch packages needed for ROCm.

7. Install ComfyUI’s requirements

Back in the ComfyUI directory, install its dependencies:

pip install -r requirements.txt

8. Run ComfyUI

Start the server:

python main.py

Then open localhost:8188 in your browser — ComfyUI is running.

A couple of things to keep in mind:

  • Ctrl+C stops the server; the browser UI will show reconnecting.
  • If you restart, always make sure the virtual environment is activated first (check with python --version).

9. Set up Flux.1-dev

To generate images with Flux.1-dev, download the models from Hugging Face and put them in the right directories:

ModelDirectory
ae.safetensorsmodels/vae
clip_l.safetensorsmodels/text_encoders
t5xxl_fp16.safetensorsmodels/text_encoders
flux1-dev.safetensorsmodels/diffusion_models

Once the files are in place, ComfyUI picks them up in the browser. You can also browse the templates in ComfyUI and search for the Flux.1-dev template to get a ready-made workflow — then just adjust the prompt and run it.

10. Generate images

Run the workflow and watch it work. On an RX 7700 XT it takes around 5 minutes to generate a Flux.1-dev image — and the results look pretty good.

Next steps

  • The ROCm nightly releases move fast — keep an eye on the RELEASES.md in the ROCm repo for new wheel commands and newly supported GPUs.
  • Flux.1-dev is a big model; if generation is too slow, experiment with smaller/faster workflows or quantized model variants — for example, Z-Image Turbo.
← All posts