Canva Craft Pitches Template Training for AI Image Generation

Daniel Okafor

Canva has released Canva Craft, its first image-generation model; the company says it trained the system on millions of human-made design templates rather than photographs. Craft can produce images at up to 2K resolution across more than 30 aspect ratios, putting the online design platform and Adobe competitor in contention with models from Google, OpenAI, ByteDance and Meta.

Canva reports that 95% of Fortune 500 companies use its platform. That reach suggests Craft could meet a broad professional audience, although the announcement gives no general-availability date, price or access terms. According to TechCrunch, Canva’s research team developed Craft to compete with the same category of image generators.

What does template-led training change?

Canva’s distinction rests on the material it used. Its training set includes millions of human-made templates from the company’s content library: professionally designed posters, social posts, presentations, infographics, menus and typography-rich layouts.

In a press release, Canva says: “While most image models are trained primarily on photographs, Craft’s training places a much greater emphasis on design.” The company adds: “Unlike flat images, these templates teach the model exactly how layout, hierarchy, and typography fit together.”

That claim shifts the unit of learning from a finished picture to an organised visual hierarchy. Canva argues that templates encode principles behind effective design, not merely the appearance of completed work. Historically, Craft enters a lineage of applied graphics—posters, menus, presentations and infographics—in which type, scale and placement carry meaning as much as pictorial realism.

This lineage also concentrates responsibility. Millions of templates give the system a large body of compositional examples, but the cited announcement does not identify their creators, licensing terms, consent arrangements or compensation.

“Millions” establishes scale, not provenance. Without those details, outsiders cannot judge how much of Craft’s apparent design intelligence comes from the structure of the library, its size or the model itself.

How does Craft generate an image?

Canva says Craft works in two steps. First, the model analyses a prompt and maps the composition:

• the placement and relative size of objects;
• the visual hierarchy; and
• the text.

The second step produces an image at up to 2K resolution in more than 30 aspect ratios, formats intended to suit different platforms. Canva says this sequence helps create more accurate details and readable text while letting users edit specific parts of an image without altering the rest. Those are consequential claims, but the announcement offers no named test set, comparison image or independent evaluation to substantiate them.

The stated output range is broad. Canva says Craft can create social-media and marketing content, realistic scenes, digital artwork and images of environments. Its no-photo-training claim does not exclude photographic-looking subjects; realism in the finished frame says nothing by itself about the source material. The company identifies no particular generated picture, however, so readers have no individual example with which to assess tone, detail, type or errors.

Can Canva prove the advantage?

The company has not disclosed Craft’s architecture, parameter count, precise template volume, safeguards against copied designs, or comparative scores against rival systems in the material described. Nor does it explain how the model handles dense layouts, changing text, small type or edits that disturb hierarchy. Those omissions leave Craft’s strongest claims as propositions rather than demonstrated advantages.

There is also a commercial question behind the technical one. Canva already operates the design platform and template library around which Craft is trained. If generation improves workflow speed without disturbing visual structure, the model could deepen the platform’s hold on design work.

If readable text and stable edits prove inconsistent, users may still need to rebuild generated compositions by hand. Canva’s claim that 95% of Fortune 500 companies use the platform gives that experiment a large audience, not a guaranteed result.

The next milestone is evidence: Craft’s template-led case will matter only if Canva can demonstrate, in live editing tests, that generated text stays readable and one change does not disturb the rest of the composition.