Canva’s new image generator does not start by making pixels. It starts by making a plan.

That is the interesting part of Canva Craft, the company’s first in-house image generation model. You type what you want, and Craft first works out the layout: what goes where, how big it should be, what matters most, and what text needs to appear. Then it renders the image.

For anyone who has asked an AI for a flyer and received beautiful nonsense, that order matters. The model was trained on millions of human-made Canva templates, including posters, social posts, presentations, infographics, menus, and typography-heavy layouts.

Why this one feels different

Most image generators learn mostly from photographs. That makes sense if you want a photo of a dog wearing sunglasses. It makes less sense if you want a sale banner with a headline, a product shot, a price, and a call to action that does not turn into alphabet soup.

Canva’s bet is that design is its own kind of knowledge. A menu is not just an image. It has hierarchy. It has spacing. It has a reason the “BURGERS” heading is bigger than the description underneath it.

Craft was built to learn that reason. Canva says the model interprets prompts into structured layouts before rendering images, which should help with readable text, object placement, and visual relationships.

And here is the part I like: Canva says users can steer, control, and edit the result more directly. Cameron Adams, Canva’s co-founder and chief product officer, put it this way: “more ways to steer, control and edit as they bring their imagination to life.”

That is the promise worth watching. Not “the AI made a picture.” The AI made a picture you can actually work with.

The detour

I have spent enough time around design tools to know the real problem is rarely the first image.

The first image is easy. It is the third one that hurts. You like the composition, but the headline is wrong. You fix the headline, and suddenly the background shifts. You fix the background, and now the logo has moved somewhere it has no business being.

That is the old loop. Generate, regret, regenerate, regret again.

Craft’s two-stage approach is meant to break that loop. Canva says the planning model handles composition first, then a separate rendering model creates the image. The company says that separation can allow more precise edits without unintentionally changing the rest of the image.

If that works as described, it changes the feel of using AI for design. You are no longer rolling dice and hoping the next roll keeps the parts you liked. You are directing a layout.

What Canva says, and what we do not know

Canva says Craft can generate images at up to 2K resolution and supports more than 30 aspect ratios, with custom dimensions also available.

The company also says prompts can be written in any language and that Craft can generate text in 55 languages.

Those are Canva’s claims, and they are worth taking seriously because the company built the model around design templates rather than photos. But claims are not the same thing as results. The useful test is simple: give it a crowded poster, ask for specific text, then change one word and see whether the rest of the design stays put.

Canva says Craft is available to users worldwide through Dream Lab and to developers through a public API. That puts it in front of a lot of people who are not AI researchers. They are people making a birthday invitation, a restaurant menu, a school presentation, or a small-business ad.

The bigger move

There is a business story underneath the creative one.

Canva has long relied on outside AI tools for parts of its platform. Building its own image model gives the company more control over quality, speed, cost, and how generation fits into its editing tools.

That is why Adams’ other quote matters. “We’re building on over a decade of understanding how designs actually come together,” he said.

He is not claiming Canva invented image generation. He is saying Canva has spent years watching how people arrange information on a page. That library of templates is the company’s advantage.

And it is a real advantage. A general-purpose model may know what a poster looks like. Canva’s model is being trained on the kinds of posters people actually make, then asking it to understand why they work.

The quiet part

There is something a little strange about an AI learning from millions of designs made by humans, then handing those humans a tool that can make more designs.

It is not automatically bad. Designers have always borrowed structures, references, and templates. The question is whether the tool gives people more control or quietly flattens everyone’s work into the same clean, competent, forgettable style.

I think the answer depends on the editing controls. If Craft lets someone change the plan, move a headline, swap a color, and keep the parts they care about, it becomes a creative instrument. If it only produces polished images and asks you to accept them, it becomes a very fast template machine.

Canva is clearly aiming for the first one. The company says the model’s design is intended to give users greater precision, consistency, and control.

That is the right goal. The next few weeks of real use will tell us whether Craft earns it.

Browse: Computer and Video Games