# Physics as Code: Inside Un-0's Oscillator-Based Image Generation > Source: > Published: 2026-06-25 23:03:36+00:00 [AI](https://www.devclubhouse.com/c/ai)Article # Physics as Code: Inside Un-0's Oscillator-Based Image Generation Unconventional AI's new model replaces neural network layers with simulated metronomes, proving physical computing can handle complex generative tasks. [Rachel Goldstein](https://www.devclubhouse.com/u/rachel_goldstein) For the past decade, the recipe for AI progress has been simple: stack more transformer layers, buy more GPUs, and burn more electricity. We are rapidly approaching the physical and economic limits of this brute-force digital scaling. The release of Un-0 by Unconventional AI points to a radically different path. Instead of relying on digital transistors to perform trillions of discrete matrix multiplications, Un-0 uses the continuous dynamics of physical systems to generate images. Un-0 is an image generator powered by a simulated system of coupled oscillators. Reaching a Fréchet Inception Distance (FID) of 6.74 on class-conditional [ImageNet](https://www.image-net.org) 64x64, the model matches the quality of leading conventional generative methods when they were first published. While the model is currently simulated on digital hardware, the long-term goal is to run these workloads on physical substrates that harness the laws of physics directly, potentially cutting energy consumption by 1,000x. For developers, Un-0 represents a shift from discrete, layer-by-layer tensor operations to continuous-time dynamical systems. Understanding how it works requires stepping away from standard feedforward architectures and looking at how physical systems self-organize. ## The Mechanics of Coupled Oscillators To understand Un-0, picture a collection of metronomes ticking on a shared, movable surface. Each metronome has a phase, which is the angle of its arm at any given moment. Because they share a surface, the vibration of one metronome subtly nudges its neighbors. Depending on the coupling strength, they will eventually synchronize into lockstep or settle into opposing rhythms. Un-0 scales this concept to thousands of simulated oscillators. These are modeled mathematically as Kuramoto oscillators, where the motion of each oscillator follows a single ordinary differential equation (ODE) applied continuously over time: $$\dot{\theta}*i = \omega_i + \sum*{j=1}^{N} K_{ij}\sin(\theta_j - \theta_i)$$ In this system, each oscillator $i$ has a phase $\theta_i$ and a natural frequency $\omega_i$. The matrix $K_{ij}$ defines the coupling strength, determining how strongly oscillator $j$ pulls or pushes oscillator $i$ toward alignment. In Un-0, the parameters to be learned during training are the coupling matrix $K$ and the natural frequencies $\omega$. By optimizing these parameters, the system is trained to guide the phases of the oscillators along specific trajectories over time. When mapped to data space, these trajectories cause coherent images to emerge from noise. ## The Simulation Paradox There is a fundamental tension at the heart of Un-0. The model is designed to prove that physical dynamical systems are highly efficient computing substrates. However, because dedicated physical oscillator hardware is not yet commercially available, Un-0 must be simulated on conventional, digital GPUs. This creates a performance paradox. Simulating continuous-time ODEs on synchronous, digital hardware is computationally expensive. GPUs are optimized for massive, parallel, discrete matrix multiplications, not for step-by-step integration of coupled differential equations. Consequently, running Un-0 on a GPU today is slower and more power-hungry than running a standard convolutional network or transformer of equivalent capacity. The current value of Un-0 is not as a drop-in replacement for Stable Diffusion on your local workstation, but as a software-level validation. It proves that physical dynamics can scale to complex generative tasks, paving the way for future analog, optical, or neuromorphic hardware that can run these equations natively at near-zero power. ## The Developer Angle: Training and Scaling Hurdles Unconventional AI has open-sourced the weights, training, evaluation, and ablation code for Un-0. For developers looking to experiment with the codebase, the workflow differs significantly from standard deep learning pipelines. Instead of backpropagating through discrete layers, training Un-0 involves backpropagating through an ODE solver, a technique similar to Neural ODEs. The training loop must solve the Kuramoto equations over a time horizon, compute the loss against the target image distribution, and update the coupling matrix $K$ and frequencies $\omega$ using frameworks like [PyTorch](https://pytorch.org). This approach introduces a major scaling bottleneck: the coupling matrix $K$. Because every oscillator can potentially couple with every other oscillator, the size of $K$ scales quadratically, $O(N^2)$, with the number of oscillators $N$. For a 64x64 image, mapping one oscillator to each pixel requires 4,096 oscillators, resulting in a coupling matrix with nearly 17 million parameters. Scaling this directly to a 512x512 image would require over 260,000 oscillators and a coupling matrix with more than 68 billion parameters. To make this practical, developers will need to explore sparse coupling matrices, localized connectivity, or latent-space variations where the oscillators operate on compressed representations rather than raw pixels. ## Placing Un-0 in the AI Timeline Using physical systems for computation is not a new idea. Un-0 builds on a long lineage of alternative computing paradigms, including Carver Mead's neuromorphic computing from 1990, Hopfield networks from 1982, and reservoir computing from the early 2000s. More recently, the research community has explored Hamiltonian networks, liquid state machines, and thermodynamic computing. What sets Un-0 apart is its scale and task complexity. Most historical work on physical substrates focused on simple classification, pattern recognition, or low-level image processing like contrast enhancement. Un-0 is the most capable image generator to date to run on a simulated physical dynamical system, proving that these unconventional architectures can handle high-dimensional generative modeling. While we are still in the early stages of this hardware-software co-design journey, Un-0 provides a concrete, open-source starting point. It shifts physical computing from a theoretical academic exercise into a practical domain that software developers can run, modify, and benchmark. ## Sources & further reading - [Un-0: Generating Images with Coupled Oscillators](https://unconv.ai/blog/introducing-un-0-generating-images-with-coupled-oscillators/)— unconv.ai - [hckr news - Hacker News sorted by time](https://hckrnews.com/)— hckrnews.com [Rachel Goldstein](https://www.devclubhouse.com/u/rachel_goldstein)· Dev Tools Editor Rachel has been embedded in the developer tooling ecosystem for nearly eight years, covering everything from IDE wars and package-manager drama to the quiet rise of AI-assisted coding. She has a soft spot for open-source maintainers and an unhealthy number of terminal emulators installed on a single laptop. ## Discussion 0 No comments yet Be the first to weigh in.