Canva on Thursday released Canva Craft, its first image-generation model built and trained in-house, giving the design platform a direct alternative to the general-purpose models it has used from other companies. Craft is available now to all Canva users and through the company’s public API, Canva said.
The model’s main difference is the order in which it works. Rather than turning a prompt straight into pixels, Craft first writes a plan for the image: which elements appear, where they sit, how large they are relative to one another, which parts should draw the eye, and what text needs to say. A second model then renders that plan at up to 2K resolution in more than 30 aspect ratios.
How Canva Craft works
Canva calls the approach a two-stage architecture. The planning stage handles composition and typography before any pixels are generated; the rendering stage turns that plan into the final image. Canva says this produces more accurate object placement and more legible text than models that generate an image in one pass.
The design focus extends to training. Canva says Craft learned from millions of human-made templates in its library, including posters, social posts, presentations, infographics and menus, rather than mainly from photographs. The company then used supervised fine-tuning and reinforcement learning, two methods for refining a model after its initial training, with feedback from its professional designers, to improve instruction following and typography. Canva says its designers compared more than 25,000 pairs of images with leading models during evaluation.
Because the output is structured, it can be converted into editable layers with Canva’s Magic Layers feature. That matters for a design tool: a generated image is more useful when a user can move a headline, replace a product, or adjust spacing without regenerating the whole scene. Canva says users can prompt Craft in any language, and the model can render text in 55 languages.
Why the design-first model matters
Canva has relied on a mix of outside image models, so an in-house model could reduce API costs and give the company more control over quality and speed, TechCrunch reported. Those costs matter as AI usage grows across a platform that serves hundreds of millions of people.
The competitive target is broader than Canva’s usual design rivals, including Adobe. Craft is aimed at the same buyers as image models from OpenAI, Google, Microsoft, ByteDance and Meta. But Canva is not trying to win every image-generation task. Its pitch is narrower: images that fit a layout, include readable text, and remain editable inside the tool where the rest of the design work happens.
Canva Craft’s early benchmark result is respectable but not dominant. Canva says Craft debuted fourth on DesignArena’s image-generation leaderboard, behind three OpenAI models, and argues that the showing makes Canva Research the second most advanced lab for image generation. The Verge reported the same fourth-place position. A top-five debut gives Canva credibility, but it also shows how crowded the front of the market remains.
Canva’s history in AI is relevant here. The company acquired Leonardo.ai in 2024, and TechCrunch reported that researchers from Leonardo worked on Craft. Canva has also added prompt-based editing and AI layers to its core product. Craft is the first time those efforts have produced a foundation model, a base model trained for many tasks, that Canva built and trained from scratch.
What to watch
Availability is broad from day one: Canva Craft is open to all Canva users in Dream Lab, and developers can call it through the public API. Canva says a Model Context Protocol server, an interface that lets AI assistants call external tools, is coming. The company is also developing faster, cheaper variants called Craft-fast and Craft-ultrafast.
The open questions are practical. Can generated images hold up when a designer needs fine print, a precise product shot or a layout that matches a brand system? Can Canva keep quality high while lowering the cost per image? And will the API and MCP integrations attract developers who would otherwise call a general-purpose model?
Canva’s safety layer, Canva Shield, covers moderation, privacy controls and reporting. Those controls are table stakes for enterprise customers, but they will not settle the quality question.
The real test is not whether Craft reaches the top of a leaderboard. It is whether a generated image arrives close enough to a finished design that a Canva user can stop switching tools. If it does, Canva will have turned its scale in templates and design workflows into an advantage that general-purpose image models cannot easily copy.
Brian Shannon is just getting his start a reporter. He attended a technical school while still in high school where he learned a variety of skills, from photography to car mechanics. Brian also helps keep Techno Secrets social media feeds up-to-date.
