Skip to content

Latest commit

 

History

History
48 lines (44 loc) · 1.67 KB

README.md

File metadata and controls

48 lines (44 loc) · 1.67 KB

Stable Diffusion

Here is the example for Stable Diffusion model conversion and inference with OpenVINO runtime.

1. Setup Environment

conda create -n stable_diffusion python=3.10
conda activate stable_diffusion
pip install -r requirements.txt --force-reinstall
sudo apt-get install git-lfs

2. Donwload Pytorch Model with GIT LFS

git clone https://huggingface.co/runwayml/stable-diffusion-v1-5

3. Convert Pytorch Model to OpenVINO Model

Convert Pytorch Model to OpenVINO FP32 Model

python ../tools/convert.py --model_id stable-diffusion-v1-5 \
    --output_dir stable-diffusion-v1-5-ov --precision FP32

Convert Pytorch Model to OpenVINO FP16 Model

python ../tools/convert.py --model_id stable-diffusion-v1-5 \
    --output_dir stable-diffusion-v1-5-ov --precision FP16

Convert Pytorch Model to OpenVINO INT8 Model with Weight Only Compression

python ../tools/convert.py --model_id stable-diffusion-v1-5 \
    --output_dir stable-diffusion-v1-5-ov --precision FP16 --compress_weights

4. Run inference with Stable Diffusion OpenVINO Model

Run Stable Diffusion OpenVINO FP32 Model on Intel CPU

python run_sd.py -c stable-diffusion-v1-5-ov/FP32 -p "A cute cat" -d CPU

Run Stable Diffusion OpenVINO FP16 Model on Intel iGPU with static shape using model cache

python run_sd.py -c stable-diffusion-v1-5-ov/FP16 -p "A cute cat" -d GPU.0 \
     --static_shape --cache_dir model_cache

Run Stable Diffusion OpenVINO INT8 Model on Intel dGPU with static shape using Model Cache

python run_sd.py -c stable-diffusion-v1-5-ov/INT8 -p "A cute cat" -d GPU.1 \
    --static_shape --cache_dir model_cache