Core ML accelerates Stable Diffusion on iPhone/iPad/Mac with 6-bit quantization

AI Impact Summary

Apple’s WWDC-enabled Core ML optimizations allow Stable Diffusion to run on-device with 6-bit palettization, reducing memory footprint while maintaining practical latency. The update relies on coremltools.optimize for model compression and introduces on-the-fly weight uncompression, plus SPLIT_EINSUM_V2 attention optimizations to improve Neural Engine throughput. Four official Stable Diffusion variants have been ported to Core ML and published to the Hugging Face Hub, with a requirement for iOS/iPadOS 17 or macOS 14 (Sonoma) developer previews to test; this enables private, offline generation and can lower cloud costs for apps relying on image synthesis.