A ComfyUI implementation of Meta's AITemplate repo for faster inference using cpp/cuda. The old AIT repo is still available for reference. You can find it here. The node in the UI is located under loaders->AIT. This new repo is behiond but is a much more stable foundation to keep AIT online. Please be patient as the repo will eventually include the same features as before.
This repo can be downloaded using ComfyUI manager.
To install the repo manually, simply git bash inside the custom_nodes
directory and use the command git clone https://github.com/FizzleDorf/ComfyUI-AIT.git
- open a terminal pathed to the current folder and use
git clone --recursive https://github.com/facebookincubator/AITemplate
. - path to
cd python
. - run
python setup.py bdist_wheel
. - If you are using a virtual environment, make sure it's activated before installing the wheel.
pip install dist/*.whl --force-reinstall
- copy and paste the
3rdparty
directory to theLib
directory of your venv or python install.
This is for compilation only, you can do the Linux install for inference only
- Extract the
ait_windows.zip
file in\ComfyUI-AIT\compile
directory. - open a cmd pathed to the current folder and use
git clone --recursive ait_windows.bundle -b fixes
. - path to
cd ait_windows/python
. - run
python setup.py bdist_wheel
. - If you cloned your ComfyUI install and you are using a virtual environment, make sure it's activated before installing the wheel.
pip install dist/*.whl --force-reinstall
for cloned comfy installs or.\python_embed\python.exe -s -m pip install
for packaged comfy.- copy and paste the
3rdparty
directory to theLib
directory of your venv or python install.
You can find already compiled modules here otherwise you can compile your own modules (requires sm80 or higher NVIDIA GPU).
a bundle for Windows compilation is included with this repo. See Windows installation instructions.
Place your downloaded or compiles Modules in the ComfyUI/models/ait directory.
While in ComfyUI, simply hook up the AIT_Unet_Loader to your workflow and select a module that is equal to or larger than the resolution you are generating at.