Un-0 generates images with simulated oscillators, betting physics beats GPUs on energy

A startup called Unconventional AI has released Un-0, an image generator that swaps the usual deep-network-on-GPU stack for a simulated system of coupled Kuramoto oscillators — the math behind metronomes that fall into sync on a shared surface. Generation starts from thousands of oscillators at random phases, a second smaller group nudges the population toward a requested ImageNet class, the system evolves under its learned coupling, and a snapshot of the final phases is decoded into pixels. Only three things are trained: the pairwise coupling matrix, each oscillator’s natural frequency, and a lightweight decoder that accounts for under 13% of parameters. On class-conditional ImageNet 64×64 it hits FID 6.74 — roughly the quality leading conventional generators showed when they first appeared.

The quality result is the hook, but the thesis is energy. The company argues the next efficiency jump requires letting physics do the computing rather than digitizing everything, and pegs the prize at around 1,000x less energy than today’s hardware. Oscillators matter here because they map cleanly onto analog substrates like CMOS, where the circuit’s own dynamics would carry out the computation instead of a GPU simulating it. Un-0 is a software simulation that validates the approach at scale before any custom silicon exists.

Un-0 lands in a long lineage of physics-inspired computing — neuromorphic, Hopfield nets, reservoir computing, thermodynamic and Kuramoto-oscillator models — and the authors claim it is the most capable image generator to date built on a simulated physical dynamical system. Weights plus training, evaluation, and ablation code are open, with the team framing it as an early, beatable baseline rather than a finished product. The open question they don’t resolve: whether the parameter-count-versus-quality curve can be pushed to the conventional frontier, and whether the promised efficiency survives the jump from simulation to real analog hardware.