I always thought running diffusion models on a phone was a joke, but it’s actually real. Off Grid is a free open-source app that uses Apple’s Core ML pipeline, directly tapping into the Neural Engine (that dedicated AI compute unit in every Apple chip since the A11) to generate images. It runs entirely locally—not a single word of your prompt gets sent to any server.
I tried it on my 15 Pro, and at 512 resolution, each image takes 8 to 15 seconds. You can watch the image gradually emerge in real time, which is pretty satisfying. The minimum requirement is an iPhone 12 with iOS 17; running a quantized model (about 1GB) takes 20 to 30 seconds per image.
If you want speed, go for the 15 Pro or later—the 8GB RAM and beefier Neural Engine make a big difference. There are two model types: quantized (around 1GB) and full precision (4GB). I’d recommend starting with the quantized version. It works offline too, perfect for killing time on the go.