Unconventional AI has introduced Un-0, an innovative image generator that utilizes a simulated system of coupled oscillators [1]. This approach has achieved a FID score of 6.74 on ImageNet 64×64, which is comparable to the quality of leading conventional image generation methods when they were first published [2]. The Un-0 model is based on the Kuramoto dynamics, a mathematical model of coupled oscillators, and does not rely on diffusion or adversarial training [3]. According to the project documentation, Un-0 "generates an image by integrating the phase dynamics of a population of coupled oscillators — no diffusion schedule, no adversary, no iterative denoising." This puts it in a distinct computational category from current dominant image generation models.

The project argues that dynamical systems of this kind can serve as computing substrates, and that mapping the approach to analog or physical hardware could yield energy efficiency gains "on the order of 1000x" compared to today's digital accelerators [4]. The Un-0 model has been released on GitHub, along with its training, evaluation, and ablation code, to facilitate community experimentation with models grounded in physical system dynamics [5]. This release is part of Unconventional AI's effort to build a new kind of computer that runs AI on the dynamics of a physical system, at a fraction of the energy today's machines need [6]. As the community continues to explore and develop unconventional AI systems, Un-0 represents a significant step toward achieving 1000x energy efficiency improvements in modern AI.

Sources

  1. Shiqi Chen, Yuhang Li, Yuntian Wang, Hanlong Chen, Aydogan Ozcan. "Optical generative models." Nature 2025.
  2. Ilker Oguz, Niyazi Ulas Dinc, Mustafa Yildirim, Junjie Ke, Innfarn Yoo, Qifei Wang, Feng Yang, Christophe Moser, Demetri Psaltis. "Optical Diffusion Models for Image Generation." NeurIPS 2024.
  3. Tiankuang Zhou, Yizhou Jiang, Zhihao Xu, Zhiwei Xue, Lu Fang. "Hundred-layer photonic deep learning." Nature Communications 2025.
  4. Jiaqi Chu, Heiner Kremer, Fabian Falck, Grace Brennan, Burcu Canakci, James Clegg, Daniel Cletheroe, Doug Kelly, Christos Gkantsidis, Michael S. Hansen, Paul Jeha, Kirill P. Kalinin, Jim Kleewein, Babak Rahmani, Saravan Rajmohan, Victor Rühle, Jannes Gladrow, Francesca Parmigiani, Hitesh Ballani. "Analog Diffusion Models." 2026.
  5. Andraž Jelinčič, Owen Lockwood, Akhil Garlapati, Guillaume Verdon, Trevor McCourt. "An efficient probabilistic hardware architecture for diffusion-like models." 2025.
  6. Stephen Whitelam. "Generative thermodynamic computing." Physical Review Letters 136, 037101, 2026.