Goal

[KIMGEONUNG]

Although previous generative prior from pretrained GANs fail to applied to general domain due to their limited representation space, recent DDPM successfully covers natural image manifold. However, their powerful generative prior usually ignore original image contents in image restoration task. In this work, we propose a method for image restoration frame work using DM(Diffusion model) repecting original contents and details in input degraded images.

[KIMGEONUNG]

Check a main assumption that naive approach fail to reconstruct original details.

• Check the results of denoise
  • If good, try more severe degradation
  • If bad, assumption is true