Diffusion models improve generation with limited group data
Two-Stage Score Estimation for Group-Specific Diffusion Models
Machine Learning
Summary
It can be hard for computers to learn how to make new things when they only have a small number of examples from a specific group. The authors developed a new way to help diffusion models, a method for generating data, by first learning what different groups share, then adjusting for each group specifically. They proved mathematically that this approach can work well even with very little data. Their tests with simulated and image data show that this two-step method can generate better results for small datasets.
What this means in practice
- •For machine learning engineers: Improve generative models for target groups with limited data by separating shared and group-specific learning components.
- •For computer vision teams: Enhance image generation quality for underrepresented categories in datasets by leveraging a two-stage diffusion modeling approach.
Authors
Huitao Yang, Oscar Hernan Madrid Padilla
Abstract
Generative modeling across related distributions raises a question: how can shared structure support accurate generation for a target distribution with limited observations? We develop a statistical framework for diffusion transfer that analyzes a shared pooled score together with a target-specific offset. Under Besov regularity of the pooled density and a stable hierarchical composition model for the offset, we derive convergence rates that separate the costs of learning these components and identify when a simpler offset yields a faster risk-bound rate than separate fitting. Under suitable boundary and regularity assumptions, we prove that the offset remains bounded and jointly regular as diffusion time approaches zero. These results allow the lower time cutoff to vanish polynomially and provide total-variation guarantees for diffusion models as density estimators. Synthetic simulations and image benchmarks provide a proof of concept for the framework.