Deep Image Prior

The paper challenges the common assumption that deep convolutional networks succeed at image generation and restoration because they learn realistic image priors from many example images. Instead, it shows that the structure of a generator network is itself sufficient to capture a great deal of low-level image statistics prior to any learning. The authors demonstrate this by using a randomly-initialized neural network as a handcrafted prior, applying it to standard inverse problems such as denoising, super-resolution, and inpainting.

Beyond these restoration tasks, the same prior can be used to invert deep neural representations for diagnosing them and to restore images from flash/no-flash input pairs, with excellent results. The approach highlights the inductive bias captured by standard generator architectures and bridges two popular families of image restoration: learning-based methods using deep convolutional networks and learning-free methods based on handcrafted priors such as self-similarity. This mattered because it reframed where a network's restoration power comes from.