Un-0: Generating Images with Coupled Oscillators
5 hours ago
- #energy-efficiency
- #image-generation
- #physical-computing
- Un-0 is an image generator powered by a simulated system of coupled oscillators, representing a physics-based computing approach.
- It achieves an FID of 6.74 on ImageNet 64×64, matching the quality of early conventional image generation methods.
- The model uses Kuramoto oscillators where coupling strengths and natural frequencies are learnable parameters, evolving via an ODE.
- Training involves learning the coupling matrix, oscillator frequencies, and a decoder that converts phases into images.
- Ablation studies show the oscillators perform non-trivial computation, improving diversity and separating classes in latent space.
- Un-0 aims for up to 1,000x energy efficiency by leveraging physical dynamics, moving away from traditional GPU-based neural networks.
- The release includes open-source model weights, training scripts, and ablation code for CIFAR-10 and ImageNet datasets.