The Diffusers app for Mac just got a serious speed boost. Version 1.1, the native counterpart to Hugging Face’s Python diffusion library, can generate images up to twice as fast depending on your hardware. That’s not marketing fluff — the team ran benchmarks across multiple Macs and found the accelerator choice makes all the difference.
The core advantage comes from Core ML. Unlike PyTorch’s mps accelerator, Core ML can spread work across the CPU, GPU, and Apple’s Neural Engine simultaneously, letting the framework decide which parts of the model run where. On a standard M1 Mac Mini, using the Neural Engine was twice as fast as the GPU. But on an M1 Max laptop, the GPU won by a wide margin — likely because that chip packs four times the GPU cores while keeping the same Neural Engine count. The team also tested running both GPU and ANE together, and surprisingly, it didn’t beat the best single-accelerator result.
Version 1.1 now automatically picks the best accelerator for your machine. That’s a meaningful quality-of-life improvement for anyone who doesn’t want to benchmark their own hardware. The update also adds guidance scale controls, a toggle to disable the safety checker, cancelable generations, model download indicators, and a seed-reuse shortcut for tweaking parameters across generations. The app stays local and private — prompts and images never leave your Mac.
The team is asking for community help filling gaps in their benchmark data, particularly for M1 Pro, M2 Pro, and M2 Max chips, which they don’t have access to. Those mid-tier configurations aren’t offered by cloud services either, so their heuristics for auto-selecting accelerators could be off. Next up: easier access to fine-tuned and Dreambooth models from the Hub, plus an iOS and iPadOS version. The Apple Silicon ecosystem has been underserved by the image generation boom so far — a native app that actually leverages the Neural Engine is a step toward fixing that.
