EVENT DOSSIER
To overcome limitations such as errors in diffusion model texts and lack of global aesthetics in code generation, the study proposes a new paradigm of “editable visual design” driven by Coding Agents. This approach uses Visual Language Models (VLM) to serve as the creative brain for demand understanding and aesthetic judgment, utilizing image generation models to build visual world simulators for synthesizing independent assets. The system employs a “think before acting” closed-loop workflow: Agents generate isolated assets, write native HTML/CSS, and iterate and optimize based on rendering feedback, ultimately delivering editable outputs with decoupled layers and real text. Users can intuitively perform mouse dragging and layout adjustments in the graphical interface. In verification scenarios such as posters and infographics, this paradigm successfully achieves both fine aesthetics and production-level editability.
Editable Visual Design 提出基于 Coding Agent 的可编辑视觉设计新范式
提出一种由 Coding Agent 驱动的新范式"Editable Visual Design"，旨在解决扩散模型文本错误及代码生成缺乏全局审美的问题。该系统将视觉语言模型（VLM）作为“创意大脑”负责需求理解与审美判断，利用图像生成模型作为“视觉世界模拟器”合成独立资产。通过“先想象后行动”的闭环工作流，Agent 生成孤立资产、编写原生 HTML/CSS 并基于渲染反馈迭代优化。最终交付具备解耦图层和真实文本的可编辑产物，支持用户在图形界面上进行直观的鼠标拖拽与布局调整。在海报和信息图等场景验证中，该范式成功实现了精细美学与生产级可编辑性。