Introducing Un-0: Generating Images with Coupled Oscillators

TL;DR

Un-0 is Unconventional AI’s first model release: an ImageNet-64² / CIFAR-10 image generator whose backbone is a simulated population of coupled Kuramoto oscillators rather than a stack of conventional neural-network layers. Phases are initialized randomly, biased by class-conditioning oscillators, integrated forward as an ODE, and decoded to pixels by a small convolutional head (~11% of parameters). Largest configuration (Un-0.n16384, 322M params, 16,384 oscillators) reaches FID 6.74 on ImageNet 64×64 — matching the quality of early-generation conventional generators (NCSN, DCGAN-TTUR, iDDPM, BigGAN) when they were first published. The bet is that a coupled-oscillator system can be implemented directly in CMOS / analog physical substrates, offering an order-of-1000× energy reduction over GPU-based digital execution; Un-0 is the software-simulation existence proof.

Key claims

The oscillator backbone follows the Kuramoto ODE \dot{\theta}_i = \omega_i + Σ_j K_{ij} sin(\theta_j - \theta_i); the only learnable parameters of the backbone are the coupling matrix K, the natural frequencies \omega, and the conventional decoder weights [Model section, Eq. 1].
Generation is a five-step pipeline: random phase init → class-oscillator coupling → forward ODE integration (explicit Euler) → snapshot phases at time T → small conv decoder to pixels; the decoder accounts for under 13% of total trainable parameters across all six released configurations [Model Architecture, Tables 1–2].
Largest ImageNet-64² model (Un-0.n16384, 16,384 oscillators, 322.44M params, 11.66% decoder) reaches FID-50k 6.74; smallest (n6656, 57.17M) reaches 8.41 [ImageNet 64×64 table].
CIFAR-10 sweep at 1024/2048/4096 oscillators reaches FID 11.01 / 9.32 / 8.76 respectively, the largest 19.43M params [CIFAR-10 table].
Training uses the drifting loss of Deng et al. 2026 with a frozen DINOv2 feature extractor and AdamW; the largest ImageNet model takes 640 B200-hours on 8×B200, with the loss computation (not the dynamics integration) being the bottleneck [Training section].
Ablation shows the trained dynamics carry the model: a decoder-only baseline (no dynamics) and a fixed-random “reservoir” (untrained Kuramoto coupling matrix) both produce significantly worse FID than trained 10-step Un-0, and increasing integration steps from 1 → 10 monotonically improves FID for trained models [Figures 4a–4b].
The dynamics-versus-decoder labor split is interpreted as: dynamics provide diversity / coverage (recall) by separating class trajectories in low-dim decoder space, decoder provides image quality (precision); the trained Kuramoto system increases peak recall over time where an untrained reservoir collapses it [§ What the dynamics are doing, Fig. 7].
The Un-0 quality–parameter Pareto frontier sits alongside or above early conventional generators (NCSN, DCGAN-TTUR, WGAN-GP, BigGAN, iDDPM, CD, TRACT) and trails newer ones (EDM, GDD); it expands the small-model frontier among compared methods [Figs. 3a/3b, Discussion].

Method

The compute engine is a population of N Kuramoto oscillators where every pair has a learned coupling strength K_{ij} and every oscillator has a learned natural frequency \omega_i. A separate smaller group of class-conditioning oscillators is coupled into the main pool via a low-rank unidirectional matrix, biasing the dynamics toward class-specific attractor regions. Inference draws random phases \theta_i ∈ [0, 2π), integrates the Kuramoto ODE forward in time T with explicit Euler, snapshots the final phase grid, and feeds it through a small conv decoder that emits pixels. Training is end-to-end: gradient signal flows from the drifting-loss-against-DINOv2-features back through the decoder, the readout time, and the ODE integration into K, \omega, and the decoder weights.

The key architectural bet is that this is not a diffusion model — there is no noising schedule, no time conditioning, no iterative denoising, no adversary. The dynamics are unguided during training (unlike diffusion / flow matching which prescribe the per-step velocity), which requires a sample-only loss. The authors choose the drifting loss of Generative Modeling via Drifting precisely because it operates on generated samples rather than per-step regression targets, and pair it with DINOv2 features as the perceptual judge.

Training scales were verified on A100, H200, and B200; the largest ImageNet model used 640 B200-hours. CIFAR-10 evaluation uses the standard reference statistics; ImageNet-64 uses the OpenAI ADM FID-50k suite.

Results

CIFAR-10 (FID-50k): n1024 → 11.01, n2048 → 9.32, n4096 → 8.76.
ImageNet 64×64 (FID-50k): n6656 → 8.41, n10240 → 8.01, n16384 → 6.74.
Pareto framing: Un-0 matches NCSN/DCGAN-TTUR/WGAN-GP/BigGAN/iDDPM/CD/TRACT-era quality; trails EDM/GDD by a clear margin; expands the small-parameter frontier among methods the authors could measure under an identical protocol.
Decoder-only ablation: without dynamics, the decoder alone produces visibly worse images and a poor FID — the decoder is not a sufficient generator on its own.
Reservoir ablation (fixed random coupling): lies between decoder-only and trained Un-0, indicating even random Kuramoto dynamics provide useful class-separable input to the decoder, but training the dynamics adds a robust further drop in FID.
Integration-step sweep (1 vs 10 steps): 1-step trained dynamics ≈ reservoir on CIFAR-10 (linearization gives no benefit); 10-step trained dynamics strictly beat reservoir; degradation when running many more steps at inference than at training is ~3%, so the model is not over-fit to the integrator.
Diversity vs quality decomposition (Fig. 7): at the largest size, trained dynamics improve FID by ~2.5 over reservoir, and the improvement is dominated by recall (+0.10 at T=1) — i.e. dynamics preserve diversity that the decoder otherwise collapses.
What’s not yet there: EDM-class quality, multi-resolution scaling beyond 64×64, and any demonstration on physical (non-simulated) oscillator hardware. The blog explicitly frames Un-0 as the “Hello World” of physics-based generative models, not a frontier-quality result.

Why it’s interesting

Un-0 is the strongest filed datapoint for an alternative compute substrate angle that this wiki has not previously tracked: a generative model whose forward pass is the integration of a physical dynamical system (Kuramoto oscillators) rather than a stack of matrix multiplications, motivated by ~1000× analog/CMOS energy reduction. It connects to the wiki at three concrete points. (1) It uses the drifting loss from Generative Modeling via Drifting (Kaiming He’s lab) — the same recipe the wiki tracks under Diffusion training efficiency — applied here to a non-neural backbone. The Un-0 training story is essentially “drifting loss + DINOv2 perceptual judge can train almost anything as a generator,” which is a useful generality claim for the drifting-models line. (2) It shares with Generative Modeling via Drifting and SoFlow: Solution Flow Models for One-Step Generative Modeling the framing that the backbone of generative modeling is a design knob (oscillators / drifting field / flow ODE), not a fixed transformer; Un-0 takes the most radical position. (3) Alex Nichol’s contemporaneous critique that Un-0 still leans on a frozen DINOv2 + a conv decoder is the right caveat — the dynamics replace the deep network but not the perceptual loss nor the readout — and the ablations honestly attribute the lift to the dynamics.