Canva’s new design model doesn’t generate images. It writes code to edit them.

Canva unveiled Canva Brush, a new vision-language model designed specifically for designers, last week. Brush is an editing tool instead of an image generator; it interprets design change requests and writes code to make those changes. It is available to Canva Pro users. Brush reads a design's content and structure, writes and executes small pieces of code to make adjustments, then verifies its results.

According to Canva, editing designs is more complex than generating them, as the model must understand the designer's intent, how the design is structured, and what needs to stay unchanged. Brush aims to solve this complexity by continually evaluating and improving its work, helping users transform a rough sketch into a polished design ready for sharing.

Canva's Brush model arrived a week after its first in-house image-generation model, Canva Craft, which launched on October 22. Brush's main challenge is editing designs, which requires understanding the designer's intentions, the design's structure, and maintaining consistency in typography, color, and spacing. The model uses an agentic loop that writes and runs code to adjust design elements and then inspects the page to guide its next step.

Canva's Brush model challenges the frontier lab behemoths by tackling complex image design editing tasks. Vladimir Beskorovainyi, an enterprise AI architect and founder of Besk Tech, argues that Canva is focusing on the right aspect by training its model on actual designs and editing tasks, rather than competing with larger models.

The key to Canva's success lies in its ability to train on the artifact (the designs) rather than the weights. Beskorovainyi emphasizes that verticalization, or focusing on a specific niche, can provide a competitive advantage when the data comes from within the product itself.

Dr. Morissa Schwartz, founder of Dr. Rissy's Writing & Marketing, suggests that Canva Brush would need to maintain the integrity of a design's elements, such as a headline in a 3-page pitch deck, while making necessary changes. She proposes testing the model by measuring unintended changes, legibility, and time to human approval, rather than relying solely on image-generation rankings.

Schwartz believes that Canva's advantage lies in its understanding of how real designers use the software, as opposed to frontier models that understand the internet's broader design trends.