2602.19523

Total: 1

#1 OSInsert: Towards High-authenticity and High-fidelity Image Composition [PDF1] [Copy] [Kimi] [REL]

Authors: Jingyuan Wang, Li Niu

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.

Subject: Computer Vision and Pattern Recognition

Publish: 2026-02-23 05:25:05 UTC