Introducing Un-0: Generating Images with Coupled Oscillators

Unconventional AI has trained a simulated network of interacting oscillators to help generate images. In Un-0, oscillators start at random phases, influence one another through learned connections, and evolve for a set duration. A conventional neural decoder converts their final state into pixels. The model report describes generation on datasets including 64-by-64-pixel ImageNet images.

The researchers also trained versions with the dynamics removed or left untrained. Their comparisons indicate that learning the oscillator behavior contributes to image generation beyond what the decoder achieves alone. This establishes a computational role for the simulated dynamics; the company's proposed energy advantage requires a physical implementation and is not measured by these experiments.