OpenVINO Stable Diffusion
stabilityai/stable-diffusion-2-1-base
This repository contains the models from stabilityai/stable-diffusion-2-1-base converted to OpenVINO, for accelerated inference on CPU or Intel GPU with OpenVINO's integration into Optimum: optimum-intel. The model weights are stored with FP16 precision, which reduces the size of the model by half.
Please check out the source model repository for more information about the model and its license.
To install the requirements for this demo, do pip install "optimum-intel[openvino, diffusers]". This installs all the necessary dependencies,
including Transformers and OpenVINO. For more detailed steps, please see this installation guide.
The simplest way to generate an image with stable diffusion takes only two lines of code, as shown below. The first line downloads the model from the Hugging Face hub (if it has not been downloaded before) and loads it; the second line generates an image.
from optimum.intel.openvino import OVStableDiffusionPipeline
stable_diffusion = OVStableDiffusionPipeline.from_pretrained("helenai/stabilityai-stable-diffusion-2-1-base-ov")
images = stable_diffusion("a random image").images
The following example code uses static shapes for even faster inference. Using larger image sizes will require more memory and take longer to generate.
If you have an 11th generation or later Intel Core processor, you can use the integrated GPU for inference, and if you have an Intel
discrete GPU, you can use that. Add the line stable_diffusion.to("GPU") before stable_diffusion.compile() in the example below.
Model loading will take some time the first time, but will be faster after that, because the model will be cached. On GPU, for stable
diffusion only static shapes are supported at the moment.
from optimum.intel.openvino.modeling_diffusion import OVStableDiffusionPipeline
batch_size = 1
num_images_per_prompt = 1
height = 256
width = 256
# load the model and reshape to static shapes for faster inference
model_id = "helenai/stabilityai-stable-diffusion-2-1-base-ov"
stable_diffusion = OVStableDiffusionPipeline.from_pretrained(model_id, compile=False)
stable_diffusion.reshape( batch_size=batch_size, height=height, width=width, num_images_per_prompt=num_images_per_prompt)
stable_diffusion.compile()
# generate image!
prompt = "a random image"
images = stable_diffusion(prompt, height=height, width=width, num_images_per_prompt=num_images_per_prompt).images
images[0].save("result.png")