Ozgur Kara, CS PhD at the Siebel School of Computing and Data Science in The Grainger College of Engineering, led a team of Google researchers with the Split-then-Merge project for successfully training a Generative AI model to composite videos without the necessary massive dataset of separated foreground and background video layers.
Written by Rudy San Miguel
Ozgur Kara
Ozgur Kara, CS PhD at the Siebel School of Computing and Data Science in The Grainger College of Engineering, led a team of Google researchers in successfully training an AI model to composite videos well without the necessary massive dataset of separated foreground and background video layers. Kara and the Biometrics AI Research team at Google worked through the summer and fall of 2026 to achieve their success with the Split-then-Merge (StM) project.
Kara’s team included: Jim Rehg, CS professor and Kara’s advisor; Du Tran, Google research scientist and advisor; Yujia Chen, Google senior machine learning engineer; Ming-Hsuan Yang, Google research scientist; and Wen-Sheng Chu, Google senior staff research scientist.
StM, a new video composition framework using Generative AI, allows for integrating a subject from one video realistically into a new background, while minding lighting, shadows and motions so that the original subject appears to belong in the new scene.
"Traditionally, composition is just about blending pixels. We decided to look at this problem through a Generative AI lens,” Kara said.
The team bypassed labeled data by amassing large collections of standard unlabeled videos and automatically splitting them into dynamic foreground and background layers. They then trained the model to merge the layers back together to reconstruct the original video.
“It’s about taking a subject and integrating it into a new background while allowing the model to make subtle adjustments—like fixing the lighting or shadows—so it actually harmonizes with the new scene dynamics. It gives us realistic results without needing manual annotation,” Kara said.
Next up for Kara is a second go-round of organizing the annual CVPR workshop hosted by AI for Creative Visual Content Generation Editing and Understanding (CVEU).
To learn more about Kara’s Split-then-Merge project, check out the StM website.
Grainger Engineering affiliations
Jim Rehg is an Illinois Grainger professor of computer science.