Generating High Fidelity Data from Low-density Regions using Diffusion Models

TL;DR

This paper introduces a modified diffusion model sampling method that effectively generates high-fidelity, novel images from low-density regions of data manifolds, addressing sample deficiency issues.

Contribution

The authors propose a novel sampling technique for diffusion models that guides generation towards low-density regions, producing high-quality, non-memorized samples.

Findings

Successfully generates high-fidelity images from low-density regions.

Guided sampling avoids memorization of training data.

Model maintains fidelity while exploring low-density areas.

Abstract

Our work focuses on addressing sample deficiency from low-density regions of data manifold in common image datasets. We leverage diffusion process based generative models to synthesize novel images from low-density regions. We observe that uniform sampling from diffusion models predominantly samples from high-density regions of the data manifold. Therefore, we modify the sampling process to guide it towards low-density regions while simultaneously maintaining the fidelity of synthetic data. We rigorously demonstrate that our process successfully generates novel high fidelity samples from low-density regions. We further examine generated samples and show that the model does not memorize low-density data and indeed learns to generate novel samples from low-density regions.