Unconventional AI, an artificial intelligence startup founded by Indian American entrepreneur and neuroscientist Naveen Rao, has released an open-source image generation model called Un-0.
Rather than relying on conventional graphics processing units (GPUs) and digital denoising steps, the model uses mathematical principles of coupled oscillators to generate visual frames.
Designed to mimic natural physical wave synchronization, the software simulation demonstrates how future analog hardware could bypass the high energy demands of modern digital computing.
Born in the United Kingdom to Indian immigrant parents, Rao grew up in Kentucky in a family of medical doctors. He studied engineering, earning a bachelor’s degree in electrical engineering and computer science from Duke University. After working for a decade as a computer architect, he completed a doctorate in computational neuroscience at Brown University.
Read: WHOOP appoints Indian American cardiologist Ami Bhatt as CMO (July 24, 2026)
Prior to launching San Francisco-based Unconventional AI, Rao co-founded Nervana Systems, which was acquired by Intel in 2016, and MosaicML, acquired by Databricks in 2023.
The release of Un-0 serves as an initial proof of concept for running artificial intelligence on non-traditional substrates. The architecture uses Kuramoto dynamics, a mathematical framework describing how rhythmic systems synchronize, to transform randomized wave phases into structured latent images.
Standard generative AI pipelines rely on heavy digital calculations across millions of transistors. By contrast, Un-0 treats dynamical physical systems as the primary computing medium. While the current release runs as a software simulation on PyTorch, the underlying design is engineered to map directly onto dedicated physical analog chips.
Read: Columbia names Pallavi Gogoi ED of Knight-Bagehot Fellowship (July 24, 2026)
Unconventional AI estimates that operating oscillator-based systems on physical analog hardware could cut computational energy consumption by up to 1,000 times compared to existing GPU accelerators.
In standardized benchmark testing, Un-0 recorded a Fréchet Inception Distance (FID) score of 6.74 on the ImageNet 64×64 dataset, achieving output quality comparable to early mainstream diffusion models. Configurations tested on the CIFAR-10 dataset achieved FID ratings as low as 8.86 using 4,096 simulated oscillators.
To support research in physical computing, Unconventional AI publicly released the model’s weights, training code, and evaluation tools under an open-source license. The codebase features parameter checkpoints ranging from 1.3 million to 322 million, integrated alongside Meta’s DINOv2 vision backbone.