Browse Teach Curate

Mask-to-Image Diffusion with Noise Injection for Output Diversity

A training-free technique that converts a sparse binary mask into many diverse yet morphologically faithful images using a pretrained controllable diffusion model. It exploits the observation that a very low-entropy conditioning input collapses diffusion outputs to near-duplicates, and that adding a small amount of artificial noise to the mask before the denoising process restores diversity without destroying the shape constraint, enabling data augmentation in data-scarce imaging domains.

Data Text

Diffusion Prism: Enhancing Diversity and Morphology Consistency in Mask-to-Image Diffusion Hao

Diffusion Prism is a training-free method for turning a sparse binary mask into many realistic, morphologically faithful images. The authors observe that when a conditioning input has very low entrop…