Editable Visual Design

Editable Visual Design. By Junyan Ye and eleven others, from Tencent Hunyuan, Sun Yat-sen University, and three other institutions. Published on the third of September. Its code repository is public and real: about thirty-five source files.

Two ways of generating visual designs exist and each fails at what the other does well. Diffusion base models are visually expressive, but their end-to-end generation yields a flattened bitmap with error-prone text, so nothing in the result can be edited afterwards layer by layer. Code-based generation through a coding agent gives you precise layout control and genuinely decoupled layers, but it lacks global aesthetic intuition and struggles to code complex visual assets. You can have the look or you can have the editability.

The paradigm proposed here is driven by a coding agent, and the division of labour is the whole idea. The vision-language model is designated the creative brain, handling requirement comprehension, task planning and aesthetic judgment. The image generation model is demoted to an on-demand visual world simulator, called only to synthesize standalone visual assets. Neither is asked to do the thing it is bad at.

The workflow is described as imagine first, then act, and it closes a loop. The agent generates isolated assets, writes native HTML and CSS around them, and then iteratively refines the design against visual rendering feedback, so the aesthetic judgment is applied to what the page actually looks like rather than to an intention. There is also an Agent Design Replay, which reproduces the creative and reasoning trajectory in the manner of a professional human designer.

What the system delivers is the part that distinguishes it: editable artifacts with decoupled layers and real text, so a user can drag things with a mouse and adjust layout in a graphical interface afterwards. Validations on posters, infographics and other scenarios are reported as achieving both refined aesthetics and production-grade editability, which are precisely the two properties that the two existing approaches could each deliver only one of.

The framing is that a bitmap is a dead end for design work regardless of how good it looks, because a design is something that gets revised. Generating the code and calling the image model only for assets keeps the output alive.

Editable Visual Design. By Junyan Ye and eleven others, from Tencent Hunyuan, Sun Yat-sen University, and three other institutions. Published on the third of September. This is the long version: the whole paper rather than the abstract.

The framing this paper reaches for is a left brain and a right brain, and the metaphor earns its place because the two failure modes really are that complementary. Neither approach to generating a visual design is a weaker version of the other. Each is missing exactly what the other has.

Start with the code side. The latest generation of models can already build structurally complete and clearly organized layouts by generating HTML, SVG and CSS directly, and that brings real engineering advantages: clean layers, interactive text, and natural support for later modification. But the authors are blunt about where it stops. Current code models are fluent in syntax, the DOM tree and Flexbox layout, and have no two-dimensional spatial sense and no visual intuition. Ask one to write layout code and you usually get the same thin, templated trio: a big headline, a card, a rounded shadow. The model knows how to write code, and not how to write code that looks good. There is a second, more concrete limitation alongside it. Code is superb at precise typography and geometric alignment and very bad at complex assets. Asking a model to hand-draw a cinematic background, a three-dimensional hero visual, natural textures or an elaborate illustration in pure code is costly and the result feels stiff, so past systems fell back on simple colour blocks, gradients or emoji as placeholders, and the output looks like an unfinished draft.

The paradigm splits the system by what each model is actually good at. The vision-language model is designated the creative brain, handling requirement comprehension, task planning and aesthetic judgment. The image generation model is demoted to an on-demand visual world simulator, called only to synthesize standalone visual assets. Neither is asked to do the thing it fails at. The workflow around them is described as imagine first, then act, and it closes a loop: the agent generates isolated assets, writes native HTML and CSS around them, and then iteratively refines the design against visual rendering feedback. That last step matters more than it sounds, because it means the aesthetic judgment is applied to what the page actually looks like once rendered, rather than to an intention held in the model's head while it writes code.

Validations on posters, infographics and other scenarios are reported as achieving both refined aesthetics and production-grade editability, which are precisely the two properties each existing approach could only deliver one of. The categories shown span campaigns, long-text design, art posters and infographics, and that range is doing work in the argument: a poster is mostly composition with little text, an infographic is mostly structured text with little composition, and a system leaning on one half of its architecture would show the seam by being much better at one than the other. Code is released.

The limitations section here is one of the most useful in the batch, and it is about the architecture rather than the numbers. The system does not create ability the underlying models lack. It routes each part of the job to whichever model is better suited, so what gets delivered is bounded by both, and if the coding agent writes weaker layout code the page is weaker however good the reference asset was. More often, the authors say, the binding constraint is the image model's sense of design: it returns something technically clean but compositionally ordinary, with no clear focal point and no colour idea to carry forward, and a reference like that gives the agent little to build on. The asset route makes its own demand too, needing the model to return a subject cleanly separated from its background. And longer pieces are harder than single ones. A multi-page deck or a full website has to stay consistent across pages, the same type scale, the same palette, a visual thread carrying from one page to the next.

That admission is the honest version of every routing architecture. Splitting a task between two specialists does not add capability, it stops each one being asked for what it cannot give. The ceiling is still the sum of what they can do, and here the authors have measured which of the two is usually the ceiling.