First Diffusion Model Running on a Game Boy: 1.08M Params, 85 Seconds Per Doodle
ready-eddy · reddit · 2026-10-30
A developer claims the first diffusion model running on a Game Boy: 16 DDIM steps on the handheld's 8 MHz CPU, 85 seconds from pure noise to doodle, no PC attached. Key specs: an MLP denoiser (784→512→512→784, 1.08M params) trained on 880k Quick, Draw! sketches; int8 weights, 4-bit activations; classifier-free guidance distilled into weights; 8.8M multiply-adds per image. Since the GB CPU can't multiply, a 24KB lookup table turns each multiply into two lookups. The 2MB ROM is almost all weights and matches the Python reference bit-for-bit. ROM and code are open-sourced.
More from Fun
- Retro-anime AI flying motorcycle art sparks QT chain on X — Loo_Atreides · 2026-10-30
- Higgsfield, Nine Months Later: The Receipts Since February, and a Lawyer's Letter — techhalla · 2026-10-30
- zetalyrae roasts MacBook Pro's bottom rounded corners with a Blindsight vampire joke — zetalyrae · 2026-10-30
- AI-generated alternate 1994: a lunar shipyard where Apollo never ended — azmackay · 2026-10-30
- bricky: a LEGO compiler built for AI agents that builds models from Python scripts — _rockt · 2026-10-30
- Robots from Asimov and Menlo Research survive brutal hardware testing — alysha_lobo · 2026-10-30