What is it about?
Cross-domain image composition aims to insert a photorealistic foreground object into a background with a different visual style. The challenge is to preserve the object's appearance and structure while achieving stylistic consistency with the background and keeping regions outside the insertion area unchanged. StyleCom uses style transfer to create precisely paired training examples, where the same object appears in both photorealistic and artistic images with a consistent spatial layout. The model takes a collage of the foreground object placed onto the background at the desired location, together with a text prompt, as guidance for generating the final composition.
Featured Image
Photo by Ilya Pavlov on Unsplash
Why is it important?
Paired examples showing the same object across photorealistic and artistic domains are difficult to obtain, making cross-domain image composition challenging to learn. StyleCom addresses this data bottleneck through a scalable approach to generating precisely paired training examples. Our experiments demonstrate improved style consistency and text alignment, together with a strong balance between foreground fidelity and background preservation. This approach could support creative image editing, visual design, and product presentation across different artistic styles.
Perspectives
For me, a central insight of this work is how carefully constructed training data and a simple image condition can work together. Content-preserving style transfer provides matched examples across visual domains, while the collage explicitly specifies the foreground object's appearance and intended location. I am very grateful to all my co-authors for their contributions and support throughout this work. I hope this work encourages further research into controllable image composition and the role of training data in preserving object identity across visual styles.
Qiang Lyu
University of the Chinese Academy of Sciences
Read the Original
This page is a summary of: StyleCom: Taming Style Transfer and Collage Prior for Training-Based Cross-Domain Image Composition, November 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3767308.3835612.
You can read the full text:
Contributors
The following have contributed to this page
