How Diffusion Prior Landscapes Shape the Posterior in Blind Deconvolution
_{Code for reproducing numerical experiments}

Spectre visualizations from FFHQ (left) and AFHQ (right).

Blind deblurring results on Kohler dataset

Requirements

• pytorch (if you're using cuda, please make sure that your cuda runtime version matches your pytorch cuda verion, since the score model need to be compiled from cpp source)
• deepinv (https://deepinv.github.io/deepinv/)

Getting started

The codebase is based on previous exellent repositories: https://github.com/yang-song/score_sde_pytorch and https://github.com/NVlabs/edm/tree/main)

• The notebook sampling_and_compute_potential.ipynb shows how to sample from diffusion models and how to evaluate the potential of a given image. It also shows that blurry images are more likely.
• The notebook eigenvalues.ipynb shows how to compute the spectra of the diffusion prior, by leveraging automatic differentiation, giving an estimation on the instrinsic dimension of the image manifold.
• The notebook blind_deblurring.ipynb shows the proposed initialization and optimization strategy for solving blind deblurring problem by mimimizing the posterior.