Boost-and-Skip: A Simple Guidance-Free Diffusion for Minority Generation

Soobin Um, Beomsu Kim, Jong Chul Ye

TL;DR: We develop a simple guidance-free approach for minority sample generation using diffusion models.

Abstract: Minority samples are underrepresented instances located in low-density regions of a data manifold, and are valuable in many generative AI applications, such as data augmentation, creative content generation, etc. Unfortunately, existing diffusion-based minority generators often rely on computationally expensive guidance dedicated for minority generation. To address this, here we present a simple yet powerful guidance-free approach called *Boost-and-Skip* for generating minority samples using diffusion models. The key advantage of our framework requires only two minimal changes to standard generative processes: (i) variance-boosted initialization and (ii) timestep skipping. We highlight that these seemingly-trivial modifications are supported by solid theoretical and empirical evidence, thereby effectively promoting emergence of underrepresented minority features. Our comprehensive experiments demonstrate that Boost-and-Skip greatly enhances the capability of generating minority samples, even rivaling guidance-based state-of-the-art approaches while requiring significantly fewer computations. Code is available at https://github.com/soobin-um/BnS.

Lay Summary:

In many AI applications—like creating new content or improving data variety—it’s important to generate rare or unusual examples, also called minority samples. These are hard to find because they don’t appear often in training data. Current methods that generate these rare samples using AI (specifically, a type called diffusion models) usually need a lot of extra computing power.

This research introduces a new and much simpler way to generate these rare examples, called Boost-and-Skip. The method only needs two small tweaks to the usual process:

- Start with more randomness (variance) to explore a wider range of possibilities.

- Skip some steps in the generation process to speed things up and better capture rare features.

Even though these changes sound small, they are backed by strong evidence and work surprisingly well. Experiments show that Boost-and-Skip can generate rare samples just as well as more complex and resource-heavy methods, but with much less computing power.

Link To Code: https://github.com/soobin-um/BnS

Primary Area: Deep Learning->Generative Models and Autoencoders

Keywords: diffusion models, minority sample generation, image generation

Submission Number: 7172