Canva has shipped its own image generation model, Canva Craft, and the obvious headline is that it now competes with OpenAI, Google, ByteDance and Meta. I don’t think that’s the interesting part. Craft sits fourth on Design Arena’s crowdsourced image leaderboard, behind three OpenAI models (GPT-Image-2.5 Sunburst, GPT-Image-2.5 Flare and GPT-Image-2). Nobody is switching tools over a fourth-place finish.
What Canva built is narrower and more useful than a general-purpose image model. I think it’s the right call.
Plan first, render second
Craft works in two stages. The first reads your prompt and produces a plan: where elements sit relative to each other, how big they are, what the visual hierarchy is, and what text needs to appear. The second stage renders the image, at up to 2K resolution and in more than 30 aspect ratios.
That split targets the two things general image models still fumble in commercial work: legible text and objects that land where you asked for them. A menu with a misspelled dish or a poster where the headline drifts into the logo is useless, however pretty the lighting is. Separating layout decisions from pixel generation is a sensible way to attack that, and it’s closer to how a human designer works anyway. You sketch the grid before you pick the photo.
The training data is the moat
Canva says it trained Craft on millions of human-made templates from its own library (posters, social posts, presentations, infographics, menus, typography-heavy layouts) instead of photographs. OpenAI and Google can’t easily copy this. They have more compute and more web imagery. They don’t have a decade of structured, layered design files made by people who were trying to sell a bake sale or a SaaS product.
Because the model learned from designs, its output can be pushed through Canva’s image-to-design conversion and split into editable layers. That’s the feature I’d actually pay attention to. A flat AI image is a dead end the moment a client asks to change the date. A layered design you can edit is a starting point. Canva has been building toward this for a while, with prompt-based editing in April and layers for AI images before that.
Researchers from Leonardo.ai, which Canva bought in 2024, contributed to the model.
Cheaper to run, and coming to agents
Canva has been renting several outside image models to power its features. Owning one gives it control over cost, quality and speed, and Canva has said as much. Craft-fast and Craft-ultrafast variants are on the way, which matters more for a product used by people churning out dozens of social posts than benchmark rank does.
The distribution plan is broad. Craft is live for all users globally today: open the three-dot menu on the left-hand side and pick Dream Lab. It’s also coming to the Canva API, and Canva plans to expose it through the Model Context Protocol so agents and third-party chatbots like ChatGPT and Gemini can call it directly. That last bit is a little cheeky, since it puts Canva’s model inside the products of the companies it’s competing with.
My one real doubt is the other half of Canva’s pitch. The company also says Craft handles realistic scenes, digital art and environmental imagery. A model trained on templates, not photos, should be strong at layouts and type. Whether it can produce a convincing photorealistic landscape is something Design Arena’s rankings don’t answer, and Canva hasn’t published numbers that would.