expectedwrong hindsight

SD3 Runs on Your Mac Now, No Big Deal

Argmax ships DiffusionKit and the gap between "frontier model" and "runs on my laptop" gets embarrassingly narrow.

2 min read 279 words #local AI #stable diffusion #apple silicon #diffusion models #MLX
hindsight — nailed it

local image generation on apple silicon became trivial. MLX-based diffusion is now one-command easy. the pattern of 'inaccessible last month, trivial this month' continued through stable diffusion, flux, and every model since.

Stable Diffusion 3 is a genuinely impressive model — MMDiT architecture, real text rendering, the kind of image quality that made people on Twitter lose their minds for a week — and as of this week you can run it locally on an Apple Silicon Mac with a Python package and about three lines of code.

DiffusionKit, from the people at Argmax, handles the conversion and runtime — MLX under the hood, ANE when it can get it, Swift or Python depending on your persuasion. The whole thing is absurdly easy, which is either exciting or alarming depending on how much of your professional identity is wrapped up in having a GPU cluster.

There's a specific feeling when something that was inaccessible last month becomes trivial this month — not a gradual slide but a cliff edge, where the answer to "how do I run SD3?" goes from "you need a beefy NVIDIA setup and probably a cloud bill" to "pip install, go." The hardware and the software finally collapse into each other and you're just left staring at your MacBook Pro like it owes you an explanation.

Stability AI is having one of the more chaotic years in recent memory — executive turnover, open weight drama, the whole thing. The model ships anyway. Someone ports it to your laptop anyway. The center doesn't hold but the weights keep flowing.

What's worth sitting with is how this affects the baseline assumption. If SD3 runs locally in June 2024, you have to ask what the frontier looks like in eighteen months and what "runs locally" means then. The answer is probably fine and probably also something we haven't thought through.