Stable Diffusion on Apple Silicon via Core ML — on-device inference with CPU/GPU/Neural Engine

AI Impact Summary

Apple's Core ML workflow brings Stable Diffusion to on-device inference on Apple Silicon, with converted Core ML weights hosted on Hugging Face and Swift/Python pathways. The guidance highlights multiple variant choices (original vs split_einsum attention) and both compiled Swift and Python inference routes, creating hardware and pipeline tradeoffs across CPU, GPU, and Neural Engine. This enables low-latency, offline image generation for macOS/iOS apps and potential App Store distribution, but success depends on macOS 13.1+ and compatible hardware, with performance and output quality varying by model variant and device.