Diffusion models analyzed through scale-space dynamics and memorization tracking

Researchers have proposed a new perspective on diffusion models, viewing them as deterministic dynamical systems indexed by noise scale rather than solely as stochastic processes. This approach allows for the study of fixed-scale denoisers as self-maps, where fixed points correspond to critical points of smoothed data density and attractors represent modes. The persistence of examples under increasing noise levels, quantified by a critical scale $\sigma_c$, can indicate memorization due to duplication, overfitting, or outliers. Experiments with models like Stable Diffusion demonstrate that this critical scale effectively tracks memorization and provides insights into image spatial distribution and caption dependence. AI

IMPACT Provides a novel analytical framework for understanding diffusion model behavior and memorization, potentially guiding future model development and evaluation.

RANK_REASON The item is an academic paper published on arXiv detailing a new theoretical framework for understanding diffusion models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• alphaXiv
• arXiv
• CatalyzeX
• CORE Recommender
• Cristina López Amado
• DagsHub
• Diffusion Models
• Gotit.pub
• Hugging Face
• IArxiv Recommender
• ScienceCast
• Stable Diffusion

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →