Un-0 generates images via coupled oscillators

Un-0 is a massive validation of physics-based computing, proving that simulated dynamical systems can handle complex generative AI tasks without traditional deep network backbones. The true value lies not in its current image quality, but in demonstrating a viable path toward 1,000x more energy-efficient analog AI hardware.

• –Instead of digital processors calculating weights, Un-0 relies on the natural self-organization of coupled Kuramoto oscillators to represent image latents.
• –Achieving a 6.74 FID on ImageNet 64x64 matches early versions of models like BigGAN and diffusion baselines, validating the approach on standard benchmarks.
• –The model weights, training, and ablation code are fully open-sourced, enabling the community to experiment with and build on top of dynamical system architectures.
• –The ultimate goal is running these models on analog CMOS or other physical substrates, which could reduce inference energy consumption by up to 1,000x.