Coding agents combine generated imagery with editable web-based visual layouts

The proposed workflow uses a vision-language model to plan and critique designs, an image model to create assets, and HTML/CSS to preserve editable layers and text.

PaperIndependentcs.CVarXiv:2609.04034v1 ↗

Junyan Ye · Wei Liu · Dongzhi Jiang · Zichen Wen · HaoDong Li · Zhutao Lv · +6 more

Research Digest·5 September 2026·2 min read

Read the paper →

Ye and colleagues present Editable Visual Design, a system for producing visual designs that remain structurally editable rather than being flattened into a single image. The agent generates isolated visual assets, assembles them with native HTML/CSS, and revises the result using feedback from rendered previews; the authors report successful applications to posters, infographics, and related formats.

What they did

The authors built a closed-loop design agent around a vision-language model (VLM), which interprets requirements, plans the composition, writes layout code, and judges rendered results. An image-generation model is called when the design requires visual assets that would be difficult to create directly in code.

The resulting assets are placed in native HTML/CSS as separate elements alongside real text. The system then renders the composition, reviews it visually, and iterates under an “imagine first, then act” workflow. An additional component, Agent Design Replay, records or reproduces the agent’s design and reasoning trajectory.

Key findings

The system produces designs with decoupled elements and real text rather than a single flattened bitmap.
Users can reposition elements and adjust layouts through a graphical interface, including direct mouse dragging.
The authors report that the combined asset-generation and code-based approach works across posters, infographics, and other visual-design scenarios.
Iterative rendering feedback allows the agent to revise the coded composition based on its visual appearance.

Why it matters

The work addresses a practical trade-off in automated design: image generators can create visually rich compositions but offer little structural editability, while code-generating agents preserve structure but struggle with complex artwork and overall aesthetics. Combining the two could support human revision, reuse, and downstream production without requiring users to regenerate an entire image for each change.

Caveats

The supplied abstract does not describe the evaluation protocol, dataset, baselines, human studies, quantitative metrics, or failure rates, so claims of refined aesthetics and production-grade editability cannot be independently assessed here. It also remains unclear how reliably the system separates generated assets, handles complex typography and responsive layouts, or generalizes beyond the demonstrated design categories.

Analysis

The paper fits a growing line of hybrid creative systems that use generative models for difficult visual content while retaining a symbolic or code-based representation for control. Its central contribution is architectural rather than a new image-generation model: it assigns aesthetic planning and critique to a VLM, asset synthesis to an image model, and editable structure to web code.

A key open question is whether the iterative agent reliably improves designs under controlled evaluation, especially compared with direct image generation, template-based tools, and code-only agents. The usefulness of Agent Design Replay will also depend on whether its recorded trajectory is understandable and actionable for human designers.