Abstract
Face Restoration is the task of improving the visual quality of an image or video containing a face. There are many causes for a defective image; an out of focus camera, compression related artifacts, low-resolutions images to name a few. Restoring these images is a non-trivial task as there are an infinite number of possibilities in the output space for each image. Typically, generative models are used for this task as the model needs to "generate'' or "hallucinate'' the actual face from the defective image. However, since we are dealing with faces, the generative powers of the model need to be carefully tuned to produce plausible outputs.
In this dissertation, I discuss techniques for face restoration that leverage existing high quality images of the same person. This is a realistic setting because in many situations we know beforehand the identity of the person in the degraded image. I also explore how reference images can be used to guide the generation process to allow fine-grained control of generative models. I show that by using high-quality reference images, we can outperform non-reference-based models in terms of visual quality and usability.