Diffusion Models as Cartoonists: The Curious Case of High Density Regions
This work addresses the problem of understanding and accessing high-density regions in diffusion models for researchers and practitioners, though it is incremental in nature.
The paper investigates high-density regions in diffusion models, revealing that typical samplers miss significantly higher likelihood samples, which often appear as cartoon-like or blurry images, even in datasets without such examples, and introduces a practical sampler that generates images with higher likelihood.
We investigate what kind of images lie in the high-density regions of diffusion models. We introduce a theoretical mode-tracking process capable of pinpointing the exact mode of the denoising distribution, and we propose a practical high-density sampler that consistently generates images of higher likelihood than usual samplers. Our empirical findings reveal the existence of significantly higher likelihood samples that typical samplers do not produce, often manifesting as cartoon-like drawings or blurry images depending on the noise level. Curiously, these patterns emerge in datasets devoid of such examples. We also present a novel approach to track sample likelihoods in diffusion SDEs, which remarkably incurs no additional computational cost. Code is available at https://github.com/Aalto-QuML/high-density-diffusion.