ODEdit: Blind Face Restoration through Ordinary Differential Equations
Primary Area: generative models
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: Image Restoration, Zero-Shot, Generative Models, Diffusion Models, Unsupervised Learning, Transfer Learning, Image Editing.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: We propose a method to restore degraded images without any knowledge about the possible degradation, solely relying on a generative prior.
Abstract: We introduce ODEdit, an unsupervised blind face restoration method. ODEdit operates without necessitating any assumptions about the nature of the degradation affecting the images and still surpasses current approaches in versatility. It is characterized by its utilization of the generative prior encapsulated within a pre-trained diffusion model, obviating the necessity for any additional fine-tuning or any handcrafted loss function. We leverage Ordinary Differential Equations for image inversion and implement a principled enhancing approach based on score-based updates to augment the realism of the reconstructed images. Empirical evaluations on face restoration reveal the robustness and adaptability of our methodology against a varied spectrum of corruption and noise scenarios. We further show how our approach synergise with other latent-based methods to outperform the state-of-the-art Blind Face Restoration methods in our experiments.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 5694
Loading