10 minutes to run stable diffusion from scratch on Windows

Downloading stable diffusion and related environment online can be very time-consuming. This note is intended to address this issue for quick testing on new device. If you want to test the latest resources (e.g. latest diffuser, onnxruntime etc.), it will not be a good reference.

1. Download all files

   File Name	Details
   Olive_SD1.5.zip	onnx fp16 models for sd v1.5
   sdDML1.5.tar.gz	packed conda enviroment including onxxruntime directml 1.15
   test_ml.py	a python scirpt to run sd model by using DirectML as backend
   latents_fp16.npy	latents data to get the same output image each time
   Anaconda3-2019.10-Windows-x86_64.exe	anaconda app

2. Click Anaconda3***.exe to install conda environment

3. Unzip sdDML1.5.tar.gz

4. Open anaconda prompt command and go to folder sdDML1.5

5. Run the commands below
   .\Scripts\activate.bat

   .\ Scripts \conda-unpack.exe

   conda list

   Then, you can double check some key packages used for stable diffusion as below

   diffusers                 0.16.1                   pypi_0    pypi
   olive-ai                  0.2.1                    pypi_0    pypi
   onnx                      1.13.1                   pypi_0    pypi
   onnxruntime-directml      1.15.0                   pypi_0    psypi
   

6. Go to the folder where has test_dml.py, and run the script
   python test_dml.py

   Note, please point out the path where sd model is located in the command before testing.

usage: test_dml.py [-h] [--model_dir MODEL_DIR] [--latents_path LATENTS_PATH] [--prompt PROMPT]
                   [--num_images NUM_IMAGES] [--batch_size BATCH_SIZE] [--num_inference_steps NUM_INFERENCE_STEPS]
                   [--dynamic_dims] [--enable_profiling] [--enable_intermediate_model]

optional arguments:
  -h, --help            show this help message and exit
  --model_dir MODEL_DIR
                        folder of stable diffusion
  --latents_path LATENTS_PATH
                        file of latents
  --prompt PROMPT       prompt
  --num_images NUM_IMAGES
                        Number of images to generate
  --batch_size BATCH_SIZE
                        Number of images to generate per batch
  --num_inference_steps NUM_INFERENCE_STEPS
                        Number of steps in diffusion process
  --dynamic_dims        Disable static shape optimization
  --enable_profiling    To save onnx profiling file
  --enable_intermediate_model
                        To save intermediate graph optimized by onnxruntime

7. local env table

   env Name	Details
   sdov	openvino 2022.3
   onnxov	openvino 2023.0