Brush: Canva's AI learns to understand and edit your creations

Canva introduces Brush, a vision-language model capable of editing editable designs, with tasks completed up to 69% faster.

Editing a line of text without touching the layout, changing the visual identity of an entire presentation, or adapting a design to multiple formats: these operations can now be entrusted to an artificial intelligence model specially trained to understand the structure of a graphic document. Named Brush, the new system developed by Canva Research intervenes directly on the elements of a design, while allowing the user to manually take back over the edits. An approach that prioritizes editing and work continuity over generating a new visual with every request. An AI that understands the structure of designs

Unlike image generation models, Brush is not limited to interpreting a text description to produce a visual result. This vision-language model analyzes the organization of a document, its composition, layers, typography, and graphic elements before making the requested changes.

In its technical overview of Brush, Canva explains that the system works through an agentic loop. The model writes and executes code to modify the document's components, observes the obtained result, and then decides if new adjustments are necessary. It can thus correct an isolated element, harmonize multiple pages, or change an artistic direction without systematically rebuilding the entire project.

This capability also relies on the distinction between elements to transform and those that must remain unchanged. A request concerning the color of a title, for example, should not lead to modifications of the images or the layout of other components. The goal is to preserve graphic consistency throughout the interventions.

Designs remain fully editable in Canva. Text, images, colors, and layouts can still be selected, moved, or edited manually after the model's pass. An important feature for design work that requires several steps of validation and correction. Up to 69% faster and 46% less expensive

Efficiency is the other axis of development for Brush. According to internal metrics published by Canva, the new model performs general graphic editing tasks up to 69% faster than the platform's previous agent, at a 46% lower cost per operation.

These results are based on a comparison with the system previously used by Canva AI, and not with competing solutions. The company also states that Brush equals or exceeds the quality of its old setup on the majority of evaluated tasks, with particularly favorable results for style changes and adapting designs to different formats.

The stakes go beyond mere execution speed. On the scale of a platform used to produce billions of designs, the resource consumption tied to generalist models represents a significant factor. Having a specialized system allows Canva to process more requests while reducing the cost of their execution. An open model at its core, specialized by Canva

Brush was not trained entirely from scratch. Canva researchers started from an open-weights model, whose identity was not specified, before specializing it in design and graphic editing tasks.

This additional training phase is based on examples of real edits, the expertise of Canva designers, and anonymized information from users who agreed to contribute to the improvement of Canva AI. The team also used reinforcement learning to allow the system to execute full tasks and progressively improve its decisions.

To evaluate the quality of the results, Canva Research developed a reward mechanism taking into account several criteria: visual appearance, adherence to composition principles, and compliance with instructions. These evaluations notably rely on an internal library of designs, hundreds of thousands of simulated interactions, and feedback from professional designers.

This specialization distinguishes Brush from generalist models capable of producing images or text, but which do not necessarily have a sufficient understanding of the constraints inherent to editing a structured document. From presentation to social content, without starting from scratch

In Canva AI, Brush can intervene on an existing design or accompany a project from a blank page. Highlighted use cases include modifying a visual mood, harmonizing several slides, or transforming a presentation into social media content.

A user can, for example, request to modify all the titles in a presentation without manually intervening on each page. They can also explore a new color palette or adapt a horizontal document to a vertical composition, then take back control of the elements directly in the editor.

The model is available to Canva Pro subscribers for supported editing tasks. It can be selected from the model menu in Canva AI, after opening or linking a design. Broader availability is planned, with no specific timeline announced.

Certain features remain in development, notably support for the Brand Kit, the use of attached documents as references, and improved handling of compositions containing numerous images. Two models for two creation steps

This development comes one week after Canva Craft, the company's first in-house image generation model. The two systems meet complementary needs: Craft produces images, while Brush intervenes on the organization and editing of the designs that use them.

They are developed by Canva Research, the company's internal lab, which brings together over 160 artificial intelligence specialists. The strategy consists of combining generalist models provided by partners with systems specialized in the platform's creative use cases.

With Brush, attention thus shifts toward a step that is sometimes less visible than initial generation: the successive corrections, variations, and adjustments that accompany the creation of a graphic document. The communicated performance metrics remain from Canva's internal evaluations, but the model's operation illustrates an evolution in AI-assisted design tools, now capable of intervening directly on the editable structure of a project rather than on its final rendering alone.