unconv-ai/Un-0

Un-0: an image generator powered by a simulated system of coupled oscillators, an example of an emerging physical computing substrate.

What it solves

Un-0 is an image-generation model that avoids the computationally expensive iterative denoising and diffusion schedules common in modern AI image generators. It aims to provide a path toward significantly lower energy consumption for AI by utilizing dynamical systems as a computing substrate, which can be mapped onto analog and physical hardware.

How it works

Instead of diffusion, Un-0 uses Kuramoto dynamics—integrating the phase dynamics of a population of coupled oscillators. The model generates images by evolving these oscillators and passing the result through a decoder. The project provides a PyTorch reference implementation with training pipelines for CIFAR-10 and ImageNet-64, utilizing DINOv2 features for training.

Who it’s for

Researchers and developers interested in unconventional AI architectures, dynamical systems, and energy-efficient image generation.

Highlights

  • Non-Diffusion Approach: Generates images without iterative denoising or adversarial training.
  • Energy Efficiency: Designed for potential 1000× lower energy usage when implemented on analog/physical hardware.
  • Pretrained Weights: Includes checkpoints for CIFAR-10 and ImageNet-64 with varying oscillator counts.
  • Ablation Suite: Includes a comprehensive suite to measure the impact of Kuramoto dynamics versus the decoder alone.

Related

  • Dispatch
    Un-0: Generating Images with Coupled Oscillators

    Un-0 is a generative model that uses a simulated system of coupled Kuramoto oscillators to achieve an FID of 6.74 on ImageNet 64x64, aiming to prove that physical dynamical systems can replace conventional neural network layers for extreme energy efficiency.

  • Project
    boogu-project/Boogu-Image

    An open-source unified image generation and editing model family that provides high-quality text-to-image generation and precise image editing with strong bilingual text rendering.

  • Project
    nachifur/RDDM

    A residual denoising diffusion model framework for high-quality image restoration and generation, supporting tasks like deraining and deblurring.

  • Project
    lucidrains/denoising-diffusion-pytorch

    A PyTorch implementation of Denoising Diffusion Probabilistic Models (DDPM) for generating high-quality images and 1D sequences.

  • Project
    facebookresearch/tuna-2

    Tuna-2 is a unified multimodal model that replaces complex vision encoders and VAEs with direct pixel embeddings to improve both image understanding and generation.