Canva launches Brush, an AI model for editing existing designs
What happened
Canva Research released Brush, a vision-language model built specifically for editing existing designs. Unlike image-generation models, Brush is trained on design-editing tasks: it reads the current state of a design, decides what code to write to change it, executes that code, checks the result visually, and adjusts if the output is off. The model arrived about a week after Canva Craft, Canva’s image-generation model aimed at producing new visuals from scratch.
Brush was built by a 160-person research team and is currently in limited rollout inside Canva’s AI design agent. According to Capital Brief, it runs 69% faster and costs 46% less to operate than the previous design agent it replaces.
Why it matters for designers
The distinction between generation and editing matters in practice. Most design work in professional contexts is not blank-canvas creation — it is adjusting, resizing, rebranding, or iterating on something that already exists. A model optimized for editing understands layout constraints, color consistency, and element relationships in ways a generation-first model typically does not.
The agentic loop — write code, check result, adjust — mirrors how a human designer would work through a revision task. That structure also means the model can catch its own errors before a person reviews the output, which changes the nature of the human review step from catching problems to approving outcomes.
For product designers and design system teams, the more significant signal is that Canva is building and publishing research on specialized editing models rather than relying on general-purpose vision-language systems. If that approach generalizes to other platforms or open-source contexts, it could shift what the baseline for AI-assisted design work looks like over the next 12 months.