CARVid: Content-Aware Retargeting of Videoswith On-Screen Graphics

Aditya Gunturu
Aditya Gunturu1,2,4*
Mackenzie Leake
Mackenzie Leake4
Ryo Suzuki
Ryo Suzuki2,3
Stephen DiVerdi
Stephen DiVerdi4

1University of Washington2University of Calgary3University of Colorado Boulder4Adobe Research

* Work done during an internship at Adobe Research.

Read the paper Watch video How it works View interactive gallery
CARVid retargets on-screen graphics to new aspect ratios while preserving readability and design intent. Given an input video with embedded graphics (a), CARVid decomposes the animated graphic from the video (b), then performs layout-aware segmentation on a reference frame to split meaningful elements such as , text blocks and branding (c). It then recomposes the graphic automatically, or exposes direct editing controls to rearrange, group, duplicate or remove elements (d). Footage © NBC News / Apple.

Teaser

Watch on YouTube ↗

Abstract

Editors routinely reframe landscape videos to portrait for social media, but lower thirds, title cards and scoreboards get cropped, shrunk or distorted along the way. CARVid is a content-aware retargeting tool that recovers these graphics from flat, already-published videos and recomposes them for the new aspect ratio. It combines foundation vision and segmentation models with a novel image processing algorithm for text-aware scaling, layout-aware segmentation and content-aware space removal, supporting both automatic and user-directed recomposition.

How it works

DECOMPOSE

Pull out the graphic

A vision-language model finds the graphic on a reference frame and a segmentation model tracks it into a clean animated layer, with no project files needed.

SEGMENT

Find its parts

Layout-aware segmentation splits the graphic into text blocks, logos and branding that can be edited on their own.

RECOMPOSE

Fit the new width

Text-aware scaling, content-aware space removal and layout reflow fit the graphic automatically, or the editor rearranges it directly.

Recomposition, from least to most disruptive

CARVid tries the gentlest operation first and moves on only when it cannot reach the target width with readable text.

The graphic scales down uniformly, animation and all, until its smallest text reaches the minimum legible height of 16 px. If it is still wider than the target, CARVid moves on.
OCR text and strong vertical edges are protected; only columns with neither can go. Each cut follows the animation as one coherent column path, so the graphic narrows without breaking its slide-in and slide-out.
A proposer offers candidate cuts and a critic picks the one along a boundary line, ignoring cuts through empty space. Here the first loop cuts along the left line and the second along the right line, separating the Oscars logo, the story headline and the NBC NOW logo so each part can be placed on its own.
Horizontal cuts: the critic first cuts along the longest line above the ticker, then along the line under the tab, separating the CBS Daily Report graphic into three rows
Cuts can be horizontal too. On this lower third the critic first picks the longest line, above the ticker, and marks the short line under the tab as low priority; the next loop cuts along that line, splitting the graphic into three rows.

Results

Each row shows the 16:9 input on the left and CARVid's 9:16 output on the right, at the same height.

Citation

To appear in the 39th Annual ACM Symposium on User Interface Software and Technology (UIST 2026).

@inproceedings{gunturu2026carvid,
  title = {CARVid: Content-Aware Retargeting of Videos with
           On-Screen Graphics},
  author = {Gunturu, Aditya and Leake, Mackenzie and
            Suzuki, Ryo and DiVerdi, Stephen},
  booktitle = {The 39th Annual ACM Symposium on User Interface
               Software and Technology},
  series = {UIST '26},
  year = {2026},
  publisher = {Association for Computing Machinery},
  doi = {10.1145/3830398.3830716}
}