Accelerating Stable Diffusion Inference on Intel CPUs

This covers techniques for optimizing the inference performance of Stable Diffusion models on Intel's latest Sapphire Rapids CPU environment.

Using Optimum Intel and OpenVINO, you can achieve powerful acceleration effects while making minimal modifications to existing Diffusers library code.

The key performance results and optimization methods are as follows:

• Over 2x speed improvement: The default float32 model's inference speed was 32.3 seconds, but after optimizing with bfloat16 via OpenVINO, it was reduced to 16.7 seconds.
• Easy model conversion: Using OVStableDiffusionPipeline, you can instantly convert a PyTorch model to OpenVINO format and run it.
• Resolution optimization: Additional performance gains can be achieved by reshaping the pipeline to fit a specific resolution.