Delve into Image Style Diffusion Towards Schrödinger Bridge Problem

Richard Jiang, YIJIE ZHU

Keywords: Diffusion Models, Schrödinger Bridge Problem, Score-Based Generative Modelling, Image Style Diffusion

TL;DR: In this work, we present the first study on investigating how to use diffusion models for image style diffusion towards Schrödinger Bridge Problem.

Abstract: Taking inspiration from the exceptional performances of Score-Based Generative Modeling (SGM) in image generation tasks, we introduce a novel Style-Diffusion method in this work. For the first time, we achieve flexible and efficient stylization transfer using SGM while preserving the semantic structures. With the prior distributions $p_{\theta}(v)$ obtained from encoding the source domain data samples, we employ approximate score-matching to estimate the drift of the reverse-time Stochastic Differential Equation (SDE) at arbitrary time step. By introducing Control Factor $\phi$, we have achieved controllable stylization in the output images. To improve computation speed, we re-formulate the original multi-end diffusion problem as a composite Schr\"odinger half bridge Problem, providing a new method for the diffusion evolution between more complex multiple distributions. Numerous empirical results and comparison with state-of-the-art methods demonstrate the superior performance of our approach in terms of stylization and extraordinary preservation of semantic structure.

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.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

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: 2725