Original title: Editable Visual Design
Junyan Ye, Wei Liu, Dongzhi Jiang, Zichen Wen +8 · September 2026
Most modern image generation models create flattened pictures. Because everything is merged into one flat layer, you cannot easily move objects around, and written text often looks messy. On the other hand, writing computer code gives you separated layers, but code alone struggles to make complex art look good. To solve this, researchers built Editable Visual Design. The system uses a vision-language model as a brain to plan and judge style, while an image generator creates separate visual parts on demand. Following an imagine-first-then-act loop, the agent writes code, inspects the visual output, and fixes mistakes. Finally, people can easily drag elements and edit real text on a screen. Tests on posters and infographics show the approach produces beautiful, fully editable designs.
When building a piece of furniture, you might sketch a plan, then cut wood pieces separately before assembling. If a part doesn't fit, you adjust it without starting over. In this paper, the system first imagines a design, creates isolated visual assets, and assembles them with code. It lets you tweak elements like text or images individually by checking how they look at each step.
Designers can create visuals that are both beautiful and easy to modify, saving hours of rework. It bridges the gap between image generation and practical editing for real-world projects.
The abstract doesn't specify limitations, but it is validated on scenarios like posters and infographics, so its performance on other design types is not established.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Editable Visual Design — Junyan Ye, Wei Liu, Dongzhi Jiang et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.