Gaussian Glasses

Research exploring: 'can a diffusion model restore discriminative structure in ambiguous or corrupted inputs?' We test this question on a dataset of camouflaged animals, seeing if the generative model can enhance the picture of the hidden animal.

TLDR: the results were largely unsuccessful. The task effectively transfers the burden of discrimination onto a generative model — a role in which it is inherently less capable than a purpose-built discriminative model.

Read the blog, or use the ImageDDPO package for your own experiments.