Apple has unveiled an artificial intelligence (AI) image generator optimized for running on Apple chips, called Stable Diffusion (SD). It is a solution that can reduce SD image generation time in Apple neural network engine hardware.
Arstechnica reported on the 3rd (local time) that Apple released ‘Stable Diffusion’, an open source AI image synthesis model that generates new images by inputting text.
SD, which Apple unveiled this time, is the fastest to create images on a high-end Nvidia GPU when running locally on a Windows or Linux PC. It takes about 8.7 seconds to render a 512×512 image on the RTX 3060.
In comparison, it took about 69.8 seconds to create a 512×512 image using unoptimized SD on the M1 Mac Mini.
Apple’s new SD optimization can produce a 512×512 50-step image on the M1 chip in 35 seconds, according to Apple’s benchmark on Github. The M2 does the job in 23 seconds, and Apple’s most powerful chip, the M1 Ultra, can do the same in just 9 seconds. This is a dramatic improvement that cuts creation time by almost half.
Apple’s Optimized SD Model is a Python package that converts SD models from PyTorch to Apple’s Core ML framework and can be used on macOS 13.1 and iOS 16.2 or later Macs, iPads, and iPhones with Apple chips. Optimization works on SD 1.4, 1.5 and the new 2.0.
Apple also cites privacy protection and avoiding cloud computing costs as advantages of running AI-generated models locally on a Mac or Apple device.
“All data that the user provides as input to the model remains on the user’s device, protecting end-user privacy,” Apple said. “After the initial download, users do not need an internet connection to use the model and this allows the use of local models. developers to reduce or eliminate server-related costs.”
Chan Park, cpark@aitimes.com
