Core ML-based Stable Diffusion on Apple Silicon — Python and Swift inference paths

AI Impact Summary

Apple's effort enables on-device Stable Diffusion inference on Apple Silicon through Core ML, with weights converted and hosted on the Hugging Face Hub. Developers can choose between the original and split_einsum attention variants and deploy via Python or Swift, using either compiled mlmodelc bundles for Swift or Python-based pipelines. This unlocks private, low-latency image generation on Macs (and iPads) and can reduce cloud inference costs, but requires macOS Ventura 13.1+ and careful packaging of model variants. Teams should plan for asset management across multiple model versions (v1-4, v1-5, v2-base) and align their build pipelines to ship the appropriate Core ML assets for their target platform.