Generating High Fidelity Data from Low-density Regions using Diffusion Models
This addresses data imbalance issues in image generation for applications requiring diverse datasets, though it is an incremental improvement on existing diffusion models.
The paper tackles the problem of sample deficiency from low-density regions in image datasets by modifying diffusion model sampling to guide it towards these regions while maintaining high fidelity. The result is successful generation of novel, high-fidelity samples from low-density regions without memorizing existing data.
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.