Canva’s users rarely ask for a photograph. They ask for a banner that fits a template, a product shot that matches a brand palette, an illustration that leaves room for a headline. That difference is the argument behind Craft, the image model Canva released on Wednesday.
Craft is Canva’s first in-house image generation model, trained from scratch by the company’s research team. Canva positioned it as a design tool rather than a photography tool: the outputs are meant to sit inside a layout, with the composition and restraint a page requires. Canva said the model is available inside its editor to paying subscribers, with a wider rollout to follow.
The company had been using external models for generation, the way most software firms did while the frontier labs shipped faster than anyone could build. Craft changes that arrangement. Training in-house gives Canva control over cost per image, latency and the style vocabulary its users actually reach for.
Why a design company would spend money on model training comes down to unit economics and dependency. Every generated image on a rented model carries a per-call price that rises with usage, and a freemium business with a very large free tier has to watch that number closely. Owning the model turns a variable cost into a fixed one, at least until the next training run.
Control matters for a second reason. Canva’s customers include marketing teams at large companies, and those buyers ask about licensing, indemnification and where the training data came from. Canva has not published details of Craft’s training corpus, a gap that will draw scrutiny from enterprise legal teams before any big contract is signed.
The model also has to look like Canva. Templates, brand kits, fonts and color palettes are the product’s grammar, and a generator that produces images outside that grammar creates work for the user. A model trained for design can be tuned toward flat backgrounds, clean edges and subjects that leave negative space, which is a different objective from photorealism.
The difference between a design model and a photo model shows up in small things: text rendered inside an image, a transparent background, a subject isolated on a plain field, an output that can sit in a layout without redrawing the rest of the page. General models treat those as edge cases. A design tool treats them as the job.
Serving generated images at Canva’s volume is a compute problem as much as a modelling one. The editor has a very large free user base, and every prompt answered on someone else’s model is a bill paid at retail. A model built for narrow tasks, producing a banner rather than a gallery piece, can be smaller and cheaper to run, which is the kind of advantage that shows up in gross margin rather than in a demo.
The competitive field is crowded. Adobe has Firefly, trained on licensed and public-domain material, and has pushed indemnification as a selling point. OpenAI’s image models sit inside countless tools. Midjourney and a dozen others compete on aesthetics. Canva’s advantage is distribution: the editor is where the work already happens.
Distribution is also a constraint. Canva’s users are not art directors. They are small business owners, teachers, marketers and students who want something usable in three minutes. A model that produces a beautiful, surprising image but cannot hold a layout is worse for them than a plain one that fits.
Canva has shipped AI features for several years, mostly as thin layers over other companies’ models. Craft is the first public sign that the company wants to own the layer underneath, and the announcement was framed in the language of a platform rather than a feature: research team, from scratch, foundation model.
The bet carries risk. Model quality is judged against the frontier, and a design-tuned model can look dated if the frontier moves toward better text rendering and instruction following. Training from scratch also means Canva owns the failures, including whatever artifacts and bias show up in output a supplier would otherwise have absorbed.
Canva was founded in Sydney in 2013 and grew by selling design software to people who had never opened professional tools. That audience is the reason an in-house model makes commercial sense: the company can tune output toward the requests it actually receives, millions of them a day, rather than toward the prompts a research lab imagines.
There is an argument that the model itself is not the moat. Workflow is. A designer who has built a brand kit, a folder of assets and a set of templates inside Canva will not leave because a competitor’s images are marginally better. The model exists to keep them from leaving for a tool that generates inside the same canvas.
For Canva, Craft is a statement about where the margin sits in design software. The company has spent a decade turning templates into a platform, and it now wants the pixels generated inside that platform to be its own. Whether customers notice the difference between an in-house model and a rented one is, for now, an open question.
They will notice cost, speed and whether the thing can make a banner that does not need to be fixed. Canva’s research team will be judged on those three, not on benchmark charts.
