Xinyu Hou, Xiaoming Li, Zongsheng Yue, Chen Change Loy
A physics-guided framework is proposed for any-to-any bokeh editing, independent of the source image's capture conditions.
Existing bokeh editing methods typically assume all-in-focus inputs or first reconstruct an all-in-focus image before rendering new bokeh. This pipeline can discard useful blur cues from the source and propagate reconstruction artifacts into the final edit.
AnyBokeh estimates the source blur state using a signed circle-of-confusion map and a disparity map, modeling the linear relation between them to extract a source-specific optical fingerprint. This fingerprint is transferred to the target focus and aperture setting. A generative editor conditioned on both source and target circle-of-confusion maps performs relative blur synthesis, enabling spatially adaptive deblurring, preservation, and defocus rendering. A high-fidelity synthetic dataset with accurate depth and metadata is constructed for physically supervised learning.
Experiments on real-world benchmarks show AnyBokeh achieves faithful and controllable editing across various tasks while avoiding the all-in-focus reconstruction and test-time calibration commonly required by existing approaches.