Unconventional AI · 2026-06-25 · notable
Un-0 — image generator built from coupled oscillators, not a neural net
Un-0 generates images by simulating a population of Kuramoto oscillators whose pairwise coupling strengths are the learned parameters. It hits FID 6.74 on ImageNet 64x64 with MIT weights and training code released.
Un-0 generates images from a network of coupled oscillators instead of a diffusion or GAN backbone — and the weights are public.
Key specs
| GitHub stars | 106 |
|---|---|
| Image net 64x64 fid | 6.74 |
| Cifar 10 fid | 8.86 |
| Hn points | 155 |
What is it?
Un-0 is a research image generator from Unconventional AI that replaces the usual deep neural net with a population of Kuramoto oscillators — simple math objects from physics that sync up over time. The coupling strengths between every pair of oscillators are the trainable weights. MIT-licensed weights, training scripts, and ablations all ship together on GitHub.
How does it work?
Each pixel-region maps to an oscillator. During inference, the oscillators are integrated forward in time and their phases get decoded back into image pixels. Training adjusts the pairwise coupling matrix so the steady-state phases reproduce target images. The codebase ships six checkpoints across CIFAR-10 and ImageNet 64x64, runnable on A100, H200, and B200 GPUs.
Why does it matter?
FID 6.74 on ImageNet 64x64 puts Un-0 in the same range as early diffusion models like NCSN and iDDPM — not state of the art, but proof that a non-neural dynamical system can learn to generate at the same quality bar. The maker pitches the approach as a step toward running AI on roughly 1,000x less energy, since coupled-oscillator hardware exists in optics and analog circuits.
Who is it for?
ML researchers and hardware-software co-design teams looking past pure deep-learning substrates.
Try it
git clone https://github.com/unconv-ai/Un-0