Our new #ICML2021 paper talks about a new formulation of a classical problem, where a “prior image” is used to help solve highly ill-posed imaging inverse problems. Basically, we use StyleGANs and the style-separation it provides to measure the difference between the sought-after and the prior image, and then formulate a constrained optimization problem in the latent space of the StyleGAN. There is something it it for everyone! We demonstrate good performance on a meaningful, realistic application of brain tumor progression with simulated MRI. And we have compressed sensing-type theoretical results that rely on StyleGAN properties that can be easily estimated.
