Reproduce “Spontaneous Symmetry Breaking in Generative Diffusion Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Gabriel Raya, Luca Ambrogioni
The generative dynamics of diffusion models are governed by spontaneous symmetry breaking, dividing the process into early linear and later attractor-driven phases. This theoretical insight allowed for a Gaussian late initialization scheme, which improved fast sampler performance and sample diversity in image generation.
Key insights
- Generative diffusion models exhibit spontaneous symmetry breaking (SSB), transitioning from a high-symmetric isotropic state to a low-symmetric data-equilibrium state.
- The generative process unfolds in two distinct phases: an early linear, mean-reverting phase around a central fixed-point, and a later attractor-driven phase towards the data manifold.
- A critical time exists where the central fixed-point loses stability, amplifying early fluctuations and playing a central role in determining the final sample diversity.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.