On June 25, 2026, Unconventional AI released "Un-0," a model that generates images through the time evolution of coupled oscillators. Its largest model achieved an FID of 6.74 on ImageNet 64×64, and the model weights along with training and evaluation code have been made available. However, what is actually running is a simulation on a GPU written in PyTorch—not an oscillator chip. The "1000x energy efficiency" the company touts is not derived from this measurement result; it is a target figure the company aims to achieve going forward by co-designing hardware and models.
What Un-0 advances is not the core idea of using oscillators for computation itself. Research on training coupled phase oscillators has existed for some time, with numerical experiments reported for logic operations and small-scale handwritten digit classification. This time, that dynamical system was incorporated into class-conditional image generation and scaled up to the 1,000-class ImageNet 64×64 task. What is new is that the question of whether the time evolution of oscillators can handle part of image generation has been made verifiable on a GPU simulation.
FID of 6.74 with 16,384 Oscillators
Un-0 for ImageNet 64×64 consists of three variants using 6,656, 10,240, and 16,384 oscillators respectively. The FID measured on 50,000 generated images was 8.41, 8.01, and 6.74 in that order. FID is a metric that measures the distance between the distributions of generated images and real images, with lower values indicating closer proximity to the reference data. However, since values can shift depending on differences in preprocessing and evaluation implementation, values from other papers cannot be compared unconditionally on the same footing.
The largest model has 322.44 million trainable parameters. Of these, 284.84 million belong to the oscillator dynamical system, and 37.61 million belong to the decoder that converts latent representations into pixels, with the decoder accounting for 11.66% of the total. FID improved as the number of oscillators increased. On the other hand, the rate of improvement when scaling up is slower than the best levels achieved by conventional models. The company itself explains that Un-0 has reached a quality level comparable to early image generation models such as BigGAN and iDDPM, but does not match later, higher-performing models such as EDM and GDD.
Care is also needed in reading the comparison chart. Unconventional AI separately displayed models that it re-measured itself using the same FID-50k procedure, and models for which published values could not be reproduced due to lack of code or checkpoints. The latter are reference values for which the evaluation procedure is not fully aligned. The FID of 6.74 indicates Un-0's achievement, but it is not a figure showing that it has surpassed current image generation models.
From Phase Time Evolution to Image Generation
The computational unit of Un-0 is the Kuramoto oscillator, each possessing its own natural frequency and phase. Each oscillator is influenced by other oscillators according to coupling strength, continuously changing its phase. During training, the coupling matrix between oscillators, the natural frequencies, and the weights of the conventional decoder are updated. Rather than passing sequentially through layers of a neural network, this design uses the motion of the entire coupled system as the computation.
Image generation begins from random phases. A small group of oscillators specifying the class is connected to the main system, and the whole system is run for a fixed duration according to the learned couplings. All phases at a specified time are read out as a latent representation, and finally the decoder converts this into an RGB image. Since the current code numerically integrates this continuous-time motion using the explicit Euler method, it ultimately runs as digital computation on the GPU.
Even so, ablations confirmed that the oscillator component is not merely decorative. Unconventional AI compared a decoder-only setup with the dynamical system removed, a system with randomly fixed couplings, and the trained system. At the largest scale, using the time evolution improved FID by about 2.5 and also increased recall—which indicates the spread of the generated distribution at the specified time—by 0.10. Since quality improved with 10 steps of time integration compared to a single-step approximation, the company interprets this as evidence that the learned nonlinear time evolution plays a role in preserving image diversity.
However, the final image quality does not emerge from the oscillators alone. The conventional decoder, which accounts for about 12% of all parameters, produces the pixels, and a DINOv2 feature extractor is also used in computing the training loss. When transferring to a physical chip, the energy-saving effect cannot be confirmed unless the entire system—including phase readout and decoding—is evaluated.
The 1000x Figure Is Not a Measured Value
The largest ImageNet model this time was trained using 8 NVIDIA B200 units, with total computation amounting to 640 B200 GPU-hours. This represents the training resources spent to produce the released model, not the power required for a single inference. Un-0's published materials contain neither measurements from a chip implementing the oscillators, nor a comparison of power consumption against GPUs. Therefore, the "1000x" figure cannot be treated as a performance achievement of Un-0.
In a separate technical document, the company clarifies how it measures its target. The comparison is against state-of-the-art generative models run on state-of-the-art GPUs or TPUs, measuring end-to-end joules per image or joules per token for obtaining output of equivalent quality. It does not judge based on TOPS/W of the computing unit alone, but also considers latency, throughput, chip area, and price. Moreover, the benchmark is not 2026 GPUs, but conventional technology around 2030, when the company expects to commercialize its product.
This definition is strict but reasonable. This is because lowering generation quality also makes it easier to reduce energy consumption, and measuring only the analog computation portion would omit the costs of data conversion and memory. Unconventional AI analyzes that in modern inference, reading model weights and moving data consumes more energy than the computation itself. If a physical system's own state can be made to carry out computation, reducing the number of round trips between memory and computation, substantial improvements could result. Un-0 is at the stage of having established, via GPU simulation, a model that maps part of image generation onto a dynamical system.
The Efficiency Race Only Truly Begins with a Physical Chip
Attempts to use physical phenomena for generative models have already progressed to actual hardware. An optical generative model published in Nature in 2025 used spatial light modulators and visible light to demonstrate image generation on MNIST and Fashion-MNIST. If decoding can be handled by the diffraction of light, digital iterative computation can be reduced. On the other hand, that same paper also explicitly noted that when converting images back to the digital domain, additional power and time are required for digital-to-analog conversion and imaging. The magnitude of energy savings varies depending on the application and system boundary.
The same verification awaits Un-0. How will the trained matrix coupling 16,384 oscillators to each other be laid out onto a physical circuit? How much power will the peripheral circuitry consume, including reading phases and passing them to the conventional decoder? And can quality be maintained even for higher-resolution images or text generation? The released code has made it possible to begin experiments on the model side, but it has not yet provided answers on the chip side.
What will determine the evaluation is the number of joules consumed by the entire physical system to produce a single image of equivalent quality. When Unconventional AI runs an oscillator chip and places measured values—including peripheral processing—alongside GPUs of the same generation, Un-0 will move toward becoming a demonstrated energy-efficient AI. The figure of 1000x will only gain meaning once it appears as the result of that comparison table.