Unconventional AI has unveiled an experimental image-generation model that replaces conventional calculation-heavy computing with a simulated network of physical oscillators. The startup says the underlying approach could eventually use up to 1,000 times less energy than current AI systems.
The model, called Un-0, is the company’s first proof of concept. Unconventional AI was founded by researchers including MIT associate professor Michael Carbin, Stanford assistant professor Sara Achour, former Google engineer MeeLan Lee and former Databricks AI chief Naveen Rao. The team published technical details on June 25 and released the model through GitHub.
Most computers perform calculations by rapidly switching billions of transistors between on and off states. Image generators based on diffusion models begin with visual noise and repeatedly calculate what must be removed to produce the requested image, typically running between 20 and 50 refinement steps.
Un-0 instead represents information through oscillators, physical systems that produce repeating waves. Connected oscillators naturally influence one another until their rhythms begin to align, a process described by the Kuramoto model.
To generate an image, the system assigns different oscillator phase patterns to categories such as shoes or trains. A smaller group configured with the desired pattern acts as the prompt and gradually pulls the wider network of randomly positioned oscillators toward the same arrangement.
The system then records the oscillators’ final phases as a numerical grid. A separate decoder converts that grid into pixel values to produce the image.
Unconventional AI argues that physical oscillator hardware could consume far less electricity because current would flow continuously through closed loops rather than being forced through trillions of transistor-switching operations. The existing Un-0 model does not yet deliver those savings because it simulates the oscillators on conventional hardware. The company ultimately plans to develop dedicated oscillator-based chips.
The researchers tested Un-0 using the CIFAR-10 and ImageNet 64×64 image datasets. On CIFAR-10, its Fréchet inception distance score improved from 11.01 with 1,024 oscillators to 8.76 with 4,096 oscillators. Lower scores indicate that generated images more closely resemble the reference data.
On ImageNet 64×64, the score improved from 8.41 with 6,656 oscillators to 6.74 with 16,384 oscillators. The researchers said the results were comparable to early image generators such as Google’s BigGAN and OpenAI’s iDDPM, although they cautioned that the figures should be treated as reference points rather than directly equivalent measurements.
The team acknowledged that current image generators still produce better overall quality while using their parameters more efficiently. Unconventional AI released the model weights, training code and ablation scripts so other researchers can test the approach and run their own simulations.
