77 fully local images in 15s on an iPhone 13 by moving SDXS to the Neural Engine
QVAC_Official · reddit · 2026-11-19
The QVAC team made fully local image generation 20x faster on an iPhone 13 (2021, 4GB RAM): same SDXS model (512x512), same 15 seconds, airplane mode.
- GPU (Metal): 3 images, 3.8s each
- Neural Engine: 77 images, 192ms each
The only change was where the model runs: the UNet and tiny decoder were moved to the Neural Engine via Core ML inside their fork of stable-diffusion.cpp, keeping the text encoder on GPU. No server, no network.
Open source and free: tetherto/qvac-ext-stable-diffusion.cpp PR #66.
More from Infra
- Intel clarifies Microsoft deal: EMIB packaging, Foundry custom silicon separate from Xeon — BenBajarin · 2026-11-19
- harness-to-api: turn your Codex subscription into an OpenAI-compatible endpoint, skip API bills — gethackteam · 2026-11-19
- Anthropic's paid inference revenue growing ~11x a year, extrapolation hits $5.2T by 2028 — ricklamers · 2026-11-19
- Rumor: Nvidia board rejects Jensen Huang's return,云巨头 rebellion predicted (unverified) — naval · 2026-11-19
- Weeks after Jensen Huang's reported firing, TSMC reportedly demands cash up front for orders — RachelVT42 · 2026-11-19
- 10 wild data center concepts: orbit, seabed, oil rigs, wind turbines and even swimming pools — clemnt · 2026-11-19