In this seminar, I will discuss several fundamental challenges and limitations associated with high-perceptual-quality image restoration methods, and propose practical restoration and compression schemes. Specifically, I will first examine deterministic image restoration algorithms and show why striving for high output quality while maintaining consistency with the input measurements inevitably leads to algorithmic instability and vulnerability to adversarial attacks.
Secondly, since the perceptual quality and distortion of the reconstructions are typically at odds with each other, a key challenge in image restoration is to minimize the distortion under a constraint of perfect output quality. To address this optimization problem, I will introduce a novel algorithm that leverages a rectified flow model to approximate the optimal solution.
Finally, I will present an innovative generative approach based on pre-trained diffusion models, which produces high-quality image samples along with their losslessly compressed bit-stream representations. This new generative framework seamlessly extends to a variety of tasks, including image compression, compressed image restoration, compressed image editing, and more generally, any compressed conditional generation task.