Canva Brush Uses Code to Make Design Edits More Precise

Nov 02, 2026 719 views
8-bit pixel art of 100 hands resting on computer mice, arranged in five rows of 20 against a dark background.

Canva Brush points to a practical shift in design AI: instead of making a new image from a prompt, it works on the editable components of a design already on the page. The model reads a project’s content and structure, translates a requested change into code, applies it, then inspects the result. For teams making frequent revisions, that approach could matter more than another image generation score.

Canva, the Australian design software company, introduced Brush last week for Canva Pro users. The launch followed Canva Craft, its first in-house image-generation model, which arrived on October 22. The two releases target different jobs: Craft generates images, while Brush handles edits to designs.

Editing is a test of restraint

A design edit has to change the requested element while preserving everything around it. That sounds simple until a prompt asks for a headline revision, a color adjustment or a new visual treatment in a page with carefully arranged text, images and spacing. A model that changes too much can turn a quick update into a review exercise.

The challenge gets harder across a multi-page project. Consider a request for a three-page bakery pitch deck with Art Nouveau visuals and backgrounds. The system has to plan the story, place copy, choose imagery and keep typography, color and spacing consistent as it moves from page to page. Each choice affects the others, so small errors can accumulate into a deck that feels inconsistent or strays from the brief.

Canva describes Brush as an agent that can evaluate its work as it goes. In the company’s account, the model writes and runs code to adjust design elements, examines the page and uses that view to decide what to do next. The goal is to take an idea from a rough sketch to a finished design through a sequence of edits, rather than treating each prompt as an isolated instruction.

“Bounded problems – for well-defined, specific tasks – are where small specialized models have been quietly beating large general ones for years.”

That framing comes from enterprise AI architect and Besk Tech founder Besk Tech, Vladimir Beskorovainyi, who argues that Brush shouldn’t be judged as a bid to beat general-purpose frontier models at everything. “Canva is not challenging frontier models; it is in fact declining to compete with them,” he told The New Stack. A design editor has a narrower job, with specific inputs and a more constrained set of acceptable outcomes.

The product may supply its own training signal

Brush began with an unspecified open-weights model. Canva says it post-trained the model on design and editing tasks, drawing on design expertise and anonymized insights from people who opted in to help improve Canva AI. The company also says it used large-scale reinforcement learning, allowing the model to work through full design tasks and learn from the results.

That training setup is central to Beskorovainyi’s argument about where Canva could build an advantage. A general model developer may be able to match a tool’s basic processing ability, he says, but the training examples are harder to reproduce when they come from a company’s own product. “The moat here is the artifact, not the weights,” he said, pointing to the design work itself as a source of information about how people assemble and revise projects.

It’s a useful distinction for product teams evaluating specialist models. A model’s performance isn’t only a question of architecture or size; the quality and relevance of its examples matter too. Canva’s user base and design software could give it access to patterns in real work, though the company’s description alone doesn’t establish how well Brush handles the full range of professional editing demands.

Small edits need measurable reliability

For marketers and publishers, the clearest test is whether a requested change stays local. Dr. Morissa Schwartz, founder of Dr. Rissy’s Writing & Marketing and Dr. Morissa Schwartz, told The New Stack she hadn’t yet used Brush, but would test it on an approved design. Her example: ask the model to change a headline, then check whether the price, disclaimer and layout remain intact.

Schwartz would compare Brush with a general-purpose model on the same design, tracking unintended changes, legibility and the time a person needs to approve the result. Those measures get closer to the operational value of an editing tool than image-generation rankings do. A polished output can still cost more time if a designer has to inspect every element for collateral changes.

“A beautiful edit can still be an expensive proofreading assignment.”

That caution puts Canva’s claims in context. The company says Brush is up to 69% faster than its previous design agent and costs 46% less per task. Those figures are self-reported comparisons with Canva’s own earlier agent; they don’t show how Brush performs against frontier models or across independent customer workflows.

Deep-tech venture studio The Capital Index’s co-founder and CTO, Rodion Sorokin, sees Canva competing through specialization and distribution. “Frontier models know the Internet; Canva knows what hundreds of millions of people actually do with a slide deck,” he said. His point is that a model tuned to design tasks may be more useful in its niche than a larger general system, particularly when it fits into a product people already use.

Editable output changes the workflow

Brush’s code-based approach also shapes what users get back. Canva says every output remains editable, including text, imagery, color, layout and smaller design details. That matters because a finished image file limits what a designer can adjust afterward; an editable project leaves room for human judgment and follow-up work.

Canva Research’s 160-person AI lab developed both Brush and Craft. On the Design Arena crowdsourced leaderboard, Canva says Craft debuted in fourth place, behind OpenAI’s GPT Image 2.5 Sunburst, GPT Image 2.5 Flare and GPT Image 2. That ranking concerns Craft, Canva’s image-generation model, and doesn’t benchmark Brush’s editing ability.

Canva says the first version of Brush is available now and that work continues on attachment references, image-heavy layouts and Brand Kit support. Those additions would test whether the model can account for more of the context that governs a real design. For now, the decisive measure is narrower: whether Brush can make the requested change while leaving the rest of the work ready to use.

The post Canva’s new design model doesn’t generate images. It writes code to edit them. appeared first on The New Stack.

Source: Adrian Bridgwater · thenewstack.io

Comments

Sign in to comment.
No comments yet. Be the first to comment.

Related Articles

Canva’s new design model doesn’t generate images. It writ...