Fractal Generative Models

  • 2025-02-25 14:28:34
  • Tianhong Li, Qinyi Sun, Lijie Fan, Kaiming He
  • 0

Abstract

Modularization is a cornerstone of computer science, abstracting complexfunctions into atomic building blocks. In this paper, we introduce a new levelof modularization by abstracting generative models into atomic generativemodules. Analogous to fractals in mathematics, our method constructs a new typeof generative model by recursively invoking atomic generative modules,resulting in self-similar fractal architectures that we call fractal generativemodels. As a running example, we instantiate our fractal framework usingautoregressive models as the atomic generative modules and examine it on thechallenging task of pixel-by-pixel image generation, demonstrating strongperformance in both likelihood estimation and generation quality. We hope thiswork could open a new paradigm in generative modeling and provide a fertileground for future research. Code is available athttps://github.com/LTH14/fractalgen.

 

Quick Read (beta)

loading the full paper ...