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

High density (ii) Low density (a) Real (b) BigGAN-deep (c) DDPM (d) DDPM (Ours) Figure 1. Real vs synthetic data. We compare synthetic images from different generative models with real images from the lowdensity (1.a.i) and high-density (1.a.ii) neighborhoods of the data manifold, respectively. In 1.b we show uniformly sampled images from BigGAN [4] and in 1.c we display images generated using the conventional uniform sampling process from the diffusion model (DDPM [10, 17]). While diffusion model achieves much higher diversity than GANs, uniform sampling from them rarely generates samples from low-density neighborhoods. (1.d) Our framework guides the sampling process in diffusion models to low-density regions and generates novel high fidelity instances from these regions. 1