Search papers, labs, and topics across Lattice.
The paper addresses the challenge of generative image composition, where existing methods struggle to simultaneously achieve high authenticity (compatibility with the background) and high fidelity (preservation of foreground details). They propose a two-stage approach: first, a high-authenticity method generates a plausible foreground shape conditioned on the background, and second, a high-fidelity method refines the foreground details based on the shape from the first stage. Experiments on the MureCOM dataset demonstrate the effectiveness of this two-stage strategy for improved image composition.
Finally, an image composition method that doesn't force you to choose between plausible integration and preserving foreground detail.
Generative image composition aims to regenerate the given foreground object in the background image to produce a realistic composite image. Some high-authenticity methods can adjust foreground pose/view to be compatible with background, while some high-fidelity methods can preserve the foreground details accurately. However, existing methods can hardly achieve both goals at the same time. In this work, we propose a two-stage strategy to achieve both goals. In the first stage, we use high-authenticity method to generate reasonable foreground shape, serving as the condition of high-fidelity method in the second stage. The experiments on MureCOM dataset verify the effectiveness of our two-stage strategy. The code and model have been released at https://github.com/bcmi/OSInsert-Image-Composition.
