Canva spent the past week showing off an AI model that creates images, and now it’s turning to the harder task of reworking them. On Thursday, the company introduced Brush, an in-house vision-language model built to edit designs. Because Brush reads both what’s in a design and how it’s arranged, Canva AI users can make precise changes without losing control of the final result. The company reports that Brush completes general editing tasks up to 69 percent faster than its previous design agent, at 46 percent lower cost per task.

Teaching its existing design model to edit as well as it designs might seem like the simpler path, so why did Canva decide editing deserved its own model? It contends that generating a design is only the first step, and that editing it is the harder problem. Getting AI to produce something close to what you pictured is the easy part. Getting it to land exactly where you want takes a different kind of work, like reshaping a design’s look for a new audience, resizing it for another platform, or aligning mismatched fonts and colors.

But getting the finer adjustments right isn’t the main reason Brush exists. Cost is. Earlier this year, Canva slowed its AI rollout because the economics weren’t ready, and general-purpose models sit at the center of that math. The company explained that these models have the scale and power to handle design edits, but that same power makes them expensive to run. When hundreds of millions of people create billions of designs on a platform, even a few extra cents per edit adds up quickly.

Cheaper edits only matter if they’re the right edits, and Canva acknowledges that speed is worthless if the result misses the mark. “The model needs to understand what you want to change, how the design is structured, and just as importantly, what needs to stay exactly as it is,” the company wrote in a blog post. “Otherwise, a simple edit request can result in a whole bunch of changes you didn’t ask for.” It illustrates the challenge with a request for a three-page bakery pitch deck in an Art Nouveau style. To pull it off, an AI agent has to plan a narrative across pages, write and place text, choose imagery, and keep typography, color, and spacing consistent throughout, with each choice affecting the others. Canva said it built Brush to handle that chain of connected decisions.

“Rather than treating design as a series of isolated decisions, Brush can continually evaluate and improve its work, helping you take an idea from a rough sketch to a polished design that’s ready to share,” the company explained. Behind the scenes, Brush writes and runs small pieces of code to adjust a design, then inspects the result before deciding what to do next. In theory, that cycle of acting and checking lets Brush catch how one change ripples through the rest of a design before the user has to.

Brush didn’t start from scratch. Canva Research, Canva’s 160-person AI lab, began with an unnamed leading open-weights model and trained it further on real design and editing tasks, drawing on in-house design expertise and anonymized insights from users who opted in to help improve Canva AI.

The team also relied on large-scale reinforcement learning, a training method in which a model attempts tasks and is rewarded for good results, to help Brush “work through complete design and editing tasks, learn from what worked, and continually shape its future behavior.” Canva built a scoring system that judges the model’s output on visual quality, adherence to design principles, and whether it completed the assigned task. That system draws on Canva’s design expertise, its template library, insights into how people edit, and “hundreds of thousands” of simulated interactions. The company added that professional designers on its staff set the standard for what counts as “good” design.

For now, Brush is available in beta to Canva Pro subscribers, and Canva plans to make it widely available soon. To start, users open a design in the editor or attach it in Canva AI, select “Brush” from the dropdown menu, and describe the change they want. They can restyle a design or image by adjusting its color scheme, changing the weather, or describing a different mood; update headers across a multi-page deck with a single request; or turn a presentation slide into an Instagram graphic.

Canva said it’s already working on Brand Kit support, the ability to reference attachments, and improvements to how Brush handles image-heavy layouts.

Brush’s release comes a week after Canva debuted Craft, its first image-generation model. It builds on the company’s second-generation AI platform, where users can use text or voice to describe what they want to create. Notably, Canva’s platform already offers agentic editing, but Brush likely scales that feature considerably.

Canva isn’t alone in turning to AI to help people edit their work. Adobe reached a similar conclusion last October with Firefly Image 5, arguing that creative professionals care as much about editing content as generating it. The model introduced layered, prompt-based editing that treats each object in an image as its own layer, so users can swap or adjust elements without regenerating the entire picture. The two companies have since taken different routes. Adobe’s editing tools operate on individual images, while Canva’s Brush works across complete designs, from a single line of copy to every slide in a deck. Adobe also leans on outside partners, offering third-party models such as Google’s Gemini and Black Forest Labs’ FLUX alongside its own Firefly. Canva uses partner models, too, but it’s betting that specialized models it trains itself will deliver better results at a lower cost.