Researchers at Hugging Face have introduced Editable Visual Design, a novel framework that leverages a coding agent guided by a vision-language model (VLM) to produce layered, editable designs. The approach addresses a critical gap in AI-driven graphic design: while diffusion models like GPT-Image-2 and Nano-Banana excel at visual expressiveness, they output flattened bitmaps with error-prone text, making post-editing difficult. Conversely, code-based generation offers precise control but lacks aesthetic intuition and struggles with complex assets.
The proposed system designates the VLM as the "creative brain," handling requirement comprehension, task planning, and aesthetic judgment. Meanwhile, an image generation model acts as an on-demand "visual world simulator," synthesizing standalone visual assets. Operating under an "imagine first, then act" closed-loop workflow, the agent generates isolated assets, writes native HTML/CSS, and iteratively refines the design based on visual rendering feedback.
A key feature is Agent Design Replay, which faithfully reproduces the creative and reasoning trajectory similar to that of professional human designers. The final output consists of editable artifacts with decoupled layers and real text, allowing users to perform intuitive mouse dragging and layout adjustments on a graphical user interface.
Validations on posters, infographics, and other scenarios demonstrate that this paradigm successfully achieves both refined aesthetics and production-grade editability. The project is open-sourced on GitHub, inviting further exploration and community contributions.