CARVid: Content-Aware Retargeting of Videos with On-Screen Graphics (Video Preview)

CARVid: Content-Aware Retargeting of Videos with On-Screen Graphics Aditya Gunturu, Mackenzie Leake, Ryo Suzuki, Stephen DiVerdi UIST 2026 Video Preview for a UIST 2026 paper Video editors routinely reframe landscape videos containing on-screen graphics, such as lower thirds or scoreboards, to portrait aspect ratios for sharing on social media. Existing methods often crop, shrink, or distort these graphics, reducing readability and disrupting their original design. Manual redesign is also slow and frustrating due to the difficulty of dealing with layered compositions or flattened videos with graphics that lack editable layers. We present CARVid, a content-aware retargeting tool for adapting on-screen graphics in videos to new aspect ratios while preserving readability and original design intent. CARVid leverages recent advances in foundation vision and segmentation models, together with a novel graphics-processing algorithm guided by a VLM critic, to detect graphical overlays, extract them from their underlying videos, and apply content-aware retargeting operations to adapt them to new aspect ratios. Our human-in-the-loop approach combines text-aware scaling, layout-aware horizontal and vertical segmentation, and content-aware space removal to support both automatic and user-directed retargeting of on-screen graphics. We conducted a technical evaluation using a custom dataset of 60 in-the-wild videos from popular internet sources that contain on-screen graphics. Results provide evidence of CARVid’s technical feasibility across a wide range of videos. Web:: https://programs.sigchi.org/uist/2026/program/content/262737