Guide: Running on AMD RX 6600 XT Windows - Fix for CLIP pkg_resources and DirectML

[orchid-boop]

Successfully got Automatic1111 running on AMD RX 6600 XT 8GB on Windows 11.
Here is what was needed for anyone with the same setup:

To get AUTOMATIC1111 running on your Radeon RX 6600 XT, we had to bypass several "NVIDIA-only" defaults and fix broken software links.
Here is the breakdown of why each step was necessary to make it work:

1. The Python 3.10.6 Fix
   Why: Stable Diffusion is built specifically for this version. Newer versions (3.11+) change how "wheels" (pre-packaged code) are built, which caused your "Failed to build wheel" and "Metadata generation failed" errors.
   The "PATH" Box: Checking "Add Python to PATH" was required so the Windows Terminal could find the python command at all.
2. The Setuptools 69.5.1 Downgrade
   Why: A recent update to a tool called setuptools (version 70+) removed a piece of code (pkg_resources) that the CLIP library (the part that connects text to images) absolutely requires.
   Result: By forcing it back to 69.5.1, we restored that missing piece so CLIP could finally install.
3. The AMD "DirectML" Strategy
   Why: By default, AUTOMATIC1111 looks for NVIDIA CUDA cores. Since you have a Radeon card, it would simply crash.
   The Fix: We used the DirectML fork and added --use-directml to the startup. This tells the software to use the Windows Direct Machine Learning layer, which allows AMD hardware to run AI tasks.
4. The "Repository Not Found" Workaround
   Why: The official link for the "Stability-AI" code was recently changed or moved, causing Error 128.
   The Fix: We had to manually "clone" a mirror or switch to the dev branch so the installer could find the actual AI code it needs to build the interface.
5. Critical Flags for the 6600 XT (8GB VRAM)
   In your webui-user.bat, these arguments were the "secret sauce":
   --skip-torch-cuda-test: Stops the program from panicking when it doesn't find an NVIDIA card.
   --precision full --no-half: Prevents the "Black Image" bug common on RX 6000 series cards.
   --medvram: Optimizes the 8GB of memory on your 6600 XT so you don't run out of space during a generation.

Hope this helps other AMD users. Tested and working as of March 2026.

[orchid-boop]

Update for anyone following this guide:
After getting the basic install working, generation times were still 15-20 minutes because the standard AUTOMATIC1111 is hardcoded for NVIDIA.
The real fix is to use the AMD DirectML fork:

1. Rename your current folder to stable-diffusion-old
2. Clone the AMD fork: git clone https://github.com/lshqqytiger/stable-diffusion-webui-amdgpu
3. Copy your models from the old folder to the new one
4. Edit webui-user.bat and set:
   set COMMANDLINE_ARGS=--use-directml --precision full --no-half --skip-torch-cuda-test --medvram
5. Launch via webui-user.bat — look for AMD Radeon RX 6600 XT in the terminal
   Result: GPU 60-70%, CPU 10-20%, generation time dropped from 20 minutes to ~2 minutes.
   The --medvram flag is required for 8GB VRAM cards or you will get an out of memory error."