Content-Edited Layout Generation for Graphic Design
IEEE Transactions on Multimedia · 2026
A new task, benchmark, and two-flow reference model for generating a layout while actively preparing the visual canvas it must inhabit.
01 / Research problem
The canvas is part of the problem.
Most layout generators assume that the background image has already been prepared for graphic design. In practice, raw visual assets often place salient content exactly where text, logos, and decorative elements need to go.
Content-Edited Layout Generation (CELG) treats background editing and foreground layout generation as one coupled problem. Instead of committing to an edited image first, the two outputs are allowed to co-adapt as design requirements change.
02 / Approach
Two synchronized generative flows.
The reference model connects pretrained image-editing and layout-generation components through synchronized trajectories and bidirectional cross-attention. Each branch can therefore respond to the evolving state of the other branch rather than receiving a fixed intermediate result.
The accompanying Search-Compare-Edit pipeline constructs instruction-based editing pairs for studying canvas readiness and image-layout coordination at scale.
03 / At a glance
What this project adds.
- 01
Formulates CELG: layout generation directly from raw visual assets, with editing and composition solved together.
- 02
Introduces the SCE data pipeline and InstructEdit benchmark for evaluating design-aware image editing.
- 03
Provides an initial two-flow reference model that exchanges information between image and layout trajectories.
04 / Publication status
Under review
Manuscript under review; no public preprint or code release yet.
This page describes ongoing research and does not imply acceptance or publication.