Unconventional AI unveiled Un-0, an image generation system that replaces traditional neural networks with simulated coupled oscillators. The approach treats physics itself as the computing substrate, encoding coupling strengths and oscillator frequencies as the primary learnable parameters rather than conventional weights. Unconventional AI built this as a proof-of-concept for radically different computer architectures.

Un-0 generates class-conditional images on ImageNet 64×64 achieving FID 6.74, matching quality levels of leading conventional methods when they first launched. The system starts from random oscillator phases, applies class conditioning through a separate oscillator group, then lets physics evolve the system for a fixed duration. A decoder under 13% of total parameters renders the final image from the oscillator states.

This work joins Kuramoto oscillators, neuromorphic computing, and other physical dynamical systems in exploring alternatives to GPU-dominated AI. The approach draws inspiration from brain rhythms and synchronization patterns, modeling computation after biological timing mechanisms. Un-0 scales these concepts to generative benchmarks while releasing all model weights and training code openly.

By demonstrating that physical substrates can handle modern AI workloads, Un-0 suggests hardware implementations could deliver the targeted 1,000x energy reduction compared to current methods. The release enables further experimentation with physics-grounded models.