Editable Visual Design, explained — demoting image generation to a sketch and shipping the artifact as HTML

The image model stops being the deliverable and becomes an imagination engine plus an asset factory, while a coding agent rebuilds the piece in HTML/CSS. A walk through the paper behind posters whose layers come apart and whose text is real text.

Editable Visual Design

What the paper claims

The original title is "Editable Visual Design" (arXiv:2609.04034, published 3 September 2026, by Junyan Ye and eleven co-authors).

The abstract, restated: diffusion base models such as GPT-Image-2 and Nano-Banana are remarkably expressive, but their end-to-end generation inherently produces flattened bitmaps with error-prone text, which rules out layer-wise post-editing. Code-based generation through coding agents gives the opposite — precise layout control and decoupled layers — but is held back by a lack of global aesthetic intuition and by how hard it is to code complex visual assets. The authors therefore propose Editable Visual Design. The VLM is designated the "creative brain" for requirement comprehension, task planning and aesthetic judgment, while the image generation model is used as an on-demand "visual world simulator" that synthesises standalone visual assets. Under an "imagine first, then act" closed loop, the agent generates isolated assets, writes native HTML/CSS, and iteratively refines the design against visual rendering feedback. Agent Design Replay then reproduces the creative and reasoning trajectory in the manner of a professional human designer. What arrives is an editable artifact with decoupled layers and real text, so users can drag and rearrange things in a GUI. Validations on posters, infographics and other scenarios show the paradigm achieving both refined aesthetics and production-grade editability.

Put plainly: the image model has been reassigned from "the one who makes the finished piece" to "the one who draws the rough" and "the one who makes the parts."

An analogy: the left-brained engineer and the right-brained illustrator

The introduction (§1) frames today's visual generation as two half-brains.

A coding agent is all left brain. It is fluent in syntax, the DOM tree and Flexbox, yet lacks two-dimensional spatial sense and visual intuition. Ask it to write layout code directly and back comes the thin, highly templated trio of "big headline, card, rounded shadow." In the paper's own words: the model knows how to write code, but not how to write code that looks good. A diffusion model is all right brain. It compresses centuries of artistic priors and its composition and lighting are first-rate, but what it emits is a raster image — text distorted, elements too deeply entangled to separate into layers.

The second structural point matters just as much: code is inherently weak at building complex assets (§1). HTML, SVG and CSS excel at precise typography and alignment, but hand-drawing a cinematic background or an elaborate illustration in pure code is costly and comes out stiff. Which is why earlier code generation fell back on colour blocks, gradients and emoji as placeholders, leaving output that looked like an unfinished draft.

The intuition: imagine first, then act

The core move is not to merge the halves but to divide the labour (§1). The VLM, as creative brain, coordinates requirement understanding, design planning, code construction and judgment of the result; the image model, as visual world simulator, is callable at any time to turn an abstract idea into a concrete picture fast.

The agent runs an "imagine first, then act" loop: it first calls the simulator for an imagined visual that establishes priors for composition, lighting and colour, and only then moves to asset generation, code writing, and rounds of render-and-reflect repair. In the system reported here, the creative brain is Codex driven by GPT-5.6 Sol and the visual simulator is GPT Image 2 (§2).

Why a bitmap cannot be the deliverable

The moment something is written out as an image, what exists is not a set of meaningful elements but a grid of pixels. Headline, subject and background all dissolve into numbers on the same plane. To "move just the headline," someone has to infer after the fact where it starts and stops — and nobody knows what belongs in the hole it leaves.

Text is the same story. A headline burned into pixels is no longer a string but a collection of shapes: you cannot fix its spelling or swap it for another language. What the paper refuses is making that container the final form. Image generation still gets used — just upstream.