Generative Models for Generating Images from Low-Density Regions

• GAN: High-fidelity, but poor coverage → Not suitable
• Autoregressive models: High coverage, but low fidelity → Not suitable
• Diffusion models: High-fidelity & High coverage → Suitable!
  • But, general uniform sampling leads to an imiation of real data density distribution (i.e. long-tailed distribution), where it generates samples from high density regions with a much higher probability than from low-density regions

Contribution

• Improved sampling process for diffusion models that can generate samples from low-density neighborhoods of the training data manifold is proposed.
• The proposed method is validated using three different metrics for neighborhood density and extensive comparisons with the baseline sampling process are provided.
• Authors found that despite a limited number of training images available from low-density regions, diffusion models successfully generalize in low-density regions, not just simply memorizing training samples.

Diffusion Models

• Forward Diffusion - Markov chain with Gaussian transition

• Reverse Diffusion - Markov chain with Gaussian transition

• (Uniform) Sampling

Generating Images from Low-Density Regions

• Hardness score 1 ($H$)