Don't expect everything to work. This is my fork for testing/sharing my WIP stuff before it gets refined for a push to the main repo.
Also if you would like to try out my model merges, I have a photo merge which uses 7 models and an anime merge which uses 9 models.
if you feel like sending me a tip you can do so here; https://www.buymeacoffee.com/xzuyn. although your money is probably better spent elsewhere.
I’ve been helping people setup Stable Diffusion and run it on their AMD graphics card (or CPU) on Windows. I’ve also wrote a basic UI for the diffusers library version to make it more user friendly. This guide is a consolidation of what I’ve learned and hopefully will help other people setup their PC to run Stable Diffusion too.
The intent of this UI is to get people started running Stable Diffusion on Windows. As such this UI won't be as feature rich as other UI, nor will it be as fast as running Stable Diffusion on Linux and ROCm.
Example screenshot:
Example img2img loopback (first image is the input, then loopbacked at 0.25 denoise [16 step DEIS]):
A lot of this document is based on other guides. I've listed them below:
- https://www.travelneil.com/stable-diffusion-windows-amd.html
- https://gist.github.com/harishanand95/75f4515e6187a6aa3261af6ac6f61269#file-stable_diffusion-md
- https://rentry.org/ayymd-stable-diffustion-v1_4-guide
- https://gist.github.com/averad/256c507baa3dcc9464203dc14610d674
- Amblyopius (optional cpu on vae and/or clip & lots of general SD help)
- deinferno (inpainting fix)
- uchuusen (inpainting resolution fix)
- jomrs (abandoned temp images fix)
You'll need to have a few things prepared first:
- Install Python 3.10: https://www.python.org/downloads/release/python-3108/
- NOTE: don't install the version of Python from the Microsoft Store
- make sure that you enable "add Python to environment variables/PATH" during the install
- Install Git for Windows: https://git-scm.com/download/win
- just install with the default recommended settings
- Have a huggingface.co account
NOTE: For Windows 10/11 you need to do an additional step. Go to Settings -> search for "Manage app execution aliases" -> disable the App Installer for "python.exe" and "python3.exe".
To check if they’re installed properly open up command prompt and run the following commands:
python --version
git --version
pip --version
There shouldn't be any "not recognized as an internal or external command" errors.
Start by creating a folder somewhere to store your project. I named mine stable_diff.
Download the setup.bat file and save it into your stable_diff folder: https://raw.githubusercontent.com/azuritecoin/OnnxDiffusersUI/main/setup.bat
NOTE: make sure you save the file as a .bat file and not a .bat.txt file.
Open command prompt and navigate to your stable_diff folder. Once there run the setup.bat file:
cd <path to your stable_diff folder>
.\setup.batThe setup batch file should create a virtual envrionment and install the Python packages. It will also download Python scripts from online repositories.
Activate the virtual environment:
.\virtualenv\Scripts\activate.bat
At this point you should be in your virtual environment and your prompt should have a (virtualenv) at the begining of the line. To exit the virtual environment just run deactivate at any time.
To restart the virtual environment after closing the command prompt window, cd back into the stable_diff folder and run the .\virtualenv\Scripts\activate.bat batch file again.
Login to huggingface:
huggingface-cli.exe login
Go to the tokens page of the huggingface website and copy your token.
Go back to the command prompt window. Make sure there's no characters selected (see below). Press Esc to de-select.
Right click the Title Bar -> Edit -> Paste -> Press Enter. You should be logged in at this point.
NOTE: when pasting, the command prompt will look like nothing has happened. This is normal behaviour, just press enter and it should update.
Go to https://huggingface.co/runwayml/stable-diffusion-v1-5 and accept the terms and conditions for the model.
Run the Python script to download and convert:
python convert_stable_diffusion_checkpoint_to_onnx.py --model_path="runwayml/stable-diffusion-v1-5" --output_path="model/stable_diffusion_onnx"
NOTE: This may take a while depending on your internet connection speed.
Althernatively, you could download the pre-converted version of the model using git:
git clone https://huggingface.co/runwayml/stable-diffusion-v1-5 --branch onnx --single-branch model/stable_diffusion_onnx
Run the txt2img_onnx.py script and check if any images were generated in the output folder. NOTE: some warnings may show up but it should be working as long as an output image is generated:
python txt2img_onnx.py
If an image was generated and it's not just a blank image then you're ready to generate art! You can use the txt2img_onnx.py script to input your own prompt for example:
python txt2img_onnx.py --prompt="tire swing hanging from a tree" --height=512 --width=512
Run the UI script and wait for everything to load:
python onnxUI.py
Once you see "Running on local URL:" open up your browser and go to "http://127.0.0.1:7860". You should be able to generate images using the web UI. To close the program, go back to the command prompt and hit ctrl-C.
If the model is on the hugging face website and it's using the diffusers library, then you can use the same convert script from the guide. Make sure you've accepted the terms for any new model. In this example I'll use waifu-diffusion.
python convert_stable_diffusion_checkpoint_to_onnx.py --model_path="hakurei/waifu-diffusion" --output_path="model/waifu_diffusion_onnx"
If the pretrained model is a .ckpt file, then you'll need to do a two step conversion. You first will need to convert from .ckpt to diffusers, then from diffusers to ONNX.
Download the .ckpt model of your choice and put them in your stable_diff folder. Run the first conversion script, using trinart2_step115000.ckpt in this example:
python convert_original_stable_diffusion_to_diffusers.py --checkpoint_path="./trinart2_step115000.ckpt" --dump_path="./trinart2_step115000_diffusers"
Then run the second conversion script:
python convert_stable_diffusion_checkpoint_to_onnx.py --model_path="./trinart2_step115000_diffusers" --output_path="model/trinart2_step115000_onnx"
NOTE: make sure the --dump_path in the first script and the --model_path is the same folder name.
Once you have your newly converted model, make sure the model is saved into the model folder (create one if it doesn't exist). Once you restarted the UI, select the model from the dropdown menu on the top left.
If you don't have a graphics card with enough VRAM or you only have onboard graphics, you can still run Stable Diffusion with the CPU. Simply add a --cpu-only flag to your command line:
python txt2img_onnx.py --cpu-only
If you want to update the program, download the latest setup.bat and overwrite the old one: https://raw.githubusercontent.com/azuritecoin/OnnxDiffusersUI/main/setup.bat
Then run the following:
.\setup.bat -update
NOTE: if you're updating from diffusers v0.5.1 and below, you will need to re-convert your models.
- If inpainting does not work for you, please follow these steps from deinferno to fix it.
- Within:
virtualenv\lib\site-packages\diffusers\pipelines\stable_diffusion\pipeline_onnx_stable_diffusion_inpaint_legacy.py - Find (likely on line 402):
sample=latent_model_input, timestep=np.array([t]), encoder_hidden_states=text_embeddings - Replace with:
sample=latent_model_input, timestep=np.array([t], dtype="float32"), encoder_hidden_states=text_embeddings
- Within: