Canva Craft: The AI Image Model Trained on Templates, and Why It Writes Copy You Never Asked For
Most AI image models learned what the world looks like from photographs. That is why they are good at a sunset and bad at a menu. A photograph has no hierarchy, no headline, no column of prices that has to line up. Ask a photo-trained model for a poster and you get a picture of a poster, with letters that look like letters until you read them.
On October 22, Canva shipped a model that learned from something else. Canva Craft was trained on "millions of human-made templates" from Canva's own library: posters, social posts, presentations, infographics, menus. Then Canva's designers graded its output against rival models, more than 25,000 image pairs, under one rule: "When automated evaluations and human judgments differed, we sided with the designer."
That is an unusual training diet, and it shows in two directions at once. Craft is very good at layout. It also has a habit of filling empty space with words nobody wrote. Both come from the same place, and once you see why, you know how to use it.
What Canva Craft does
Craft makes images up to 2K for design jobs: posters, flyers, menus, slides, infographics, social graphics, plus portraits, landscapes and product shots. It can put text on the image in 55 languages, and you can write the prompt in any language.
The interesting part is how it works. Craft runs in two stages. First a planning model writes a composition plan: what elements go in, where they sit, how big they are, what is most important, and what the text says. Then a second model paints the image from that plan.
If you have ever sketched a thumbnail before opening Photoshop, you already understand why that matters. The plan is the thumbnail. Text comes out readable more often because the model decided where the headline lives before it started painting around it.
It is also available to everyone. Craft sits in Dream Lab, which you reach from Canva's home page, not from inside the editor. Free accounts can use it. There is also a preview API for developers, where the model goes by craft_v1_2k.
On the public DesignArena leaderboard, Canva says Craft debuted fourth overall and first in the People & Portraits and Landscapes & Nature categories. BetaNews notes three OpenAI models rank above it, and that its rating slipped a little in its first day. Leaderboards move. Treat that as "competitive with the best," not "the best."
The template effect, both directions
Here is the part launch coverage has mostly skipped.
A template is a finished design with placeholder content. Every menu template has dish names and prices. Every event poster has a date, a venue, a tagline and a little line of small print at the bottom. Designers built them that way on purpose, so the layout looks complete before you type your own words in.
Train a model on millions of those and it learns something useful: what a complete layout looks like. Where the headline goes. How much breathing room a price list needs. That a poster usually has a call to action near the bottom.
It also learns that a layout is not finished until every slot is filled.
The first hands-on review I found bears this out. Gaurav Tiwari tested Craft on three agency briefs and found that the text he asked for came back correct every time: a poster headline and offers, eight dishes with prices, five onboarding steps. That is a real achievement. But Craft also added copy no one requested. Three of the four poster variations carried copy he never wrote, one with opening hours printed in the footer. Most of the menu variations gained a "Bistro Italiano" subtitle.
My read, and it is my read rather than anything Canva has said: that is the template showing through. The model learned that posters have opening hours, so it supplied some. For a designer, that is the most important single fact about Craft. The words you typed will probably be right. The words you did not type may still appear, and they will look perfectly believable.
Believable wrong copy is worse than garbled copy. Garbled copy gets caught. A plausible "Open daily 9 to 6" on a poster for a shop that closes at 5 sails straight to the printer.
Put this into practice: a Craft poster you can trust
This routine borrows from The Rundown's Craft tutorial and adds a proofing step built for the template habit.
1. Open Dream Lab and confirm the model. From Canva's home page, open Dream Lab. Check that the model name under the prompt box reads Canva Craft. Pick your ratio. The tutorial's test showed six options, including 3:4 and 1:1.
2. Write the brief like a designer's brief. Subject, style, a limited palette by name, then the headline, given exactly once, in quotation marks. Then say what you do not want: no other text, no logos, no labels. Because Craft plans before it paints, telling it about absence is as useful as telling it about content.
3. Leave room for the real details. Ask for a clear band at the bottom of the image with nothing in it. That is where your date, venue and call to action will go, typed by you.
4. Pick from the round. One prompt gives you up to four options. Treat them as a comp round. Choose on composition, not on the words.
5. Proof against your brief, not against the image. This is the step that matters. Make a list of every word you asked for. Then read the image and circle every word that is not on your list. Opening hours, taglines, web addresses, prices, small print. Delete or regenerate anything you cannot vouch for. Then read your requested headline letter by letter at full size and at phone size.
6. Set the facts as real type. Place the artwork in a normal Canva design and add the date, venue and call to action as text boxes. That way a client's change of date is a ten-second edit, not a new generation.
7. Use Magic Layers on a copy, if you need it. Select the image, then Edit, then Magic Layers, which tries to split the picture into editable text and separate shapes. Tiwari found it turned his menu into editable text and graphics but lost the dotted leader lines. The tutorial warns that a very detailed image (its example was a spaceship) may not separate cleanly. It is a Premium feature and draws on the same AI allowance, so run it only on the version you are keeping.
If you build things in code, the Connect API takes the same kind of prompt. Canva's developer post says to "Expect roughly 50 to 60 seconds end to end," because the planning step comes first. It is a preview, limited to 20 requests a minute per user, and Canva says apps built on preview APIs cannot be published to all Canva users. Good for internal tools, not for a product.
What it costs
Craft does not have a price tag of its own. Every generation spends from your Canva AI allowance, and Canva does not publish a per-image figure.
According to Tiwari's review, the free plan includes up to 20 standard or premium AI uses, Pro gets roughly ten times that, Business twenty times, and an AI Pass add-on stretches it further. He calls these best-case numbers and notes the Pro wording is ambiguous. Plan prices vary by country. Those multipliers are one reviewer's reading, so check Canva's pricing page for your own plan before you budget a job around them.
The practical consequence: on a free account, a poster job that takes four prompts and a Magic Layers pass eats a real share of your month. Write the brief carefully before you press the button. That is the same advice the proofing routine gives you, which is convenient.
Where it breaks
It writes copy you did not ask for. Covered above, and it bears repeating because it is the failure that reaches clients.
It does not know your brand. Tiwari found Dream Lab did not apply his Brand Kit. Describe colours and type in the prompt, then fix them in the editor.
Edits redraw the whole picture. Asking Craft to remove the invented opening hours worked in his test, but the entire image was re-rendered and small details shifted. Canva's announcement describes "targeted changes without unnecessarily altering everything else" as a design goal. In practice, expect drift.
Fewer shapes than advertised. Canva's announcement says more than 30 aspect ratios. The API reference lists six, and so did the tutorial's Dream Lab test.
Small print and technical drawings are weak. Canva says so itself and promises to improve "fine print and technical drawings." Tiwari also flagged small infographic text as hard to read on a phone.
Indemnity is narrow. Canva allows commercial use of AI output under its terms. Tiwari notes Canva's IP indemnity for AI content covers eligible Enterprise customers, not Pro or Business. If a client contract asks you to guarantee your images, that matters.
It is two days old. The leaderboard placing is early, the hands-on evidence is one careful reviewer and one tutorial, and Canva has promised faster and cheaper versions, Craft-fast and Craft-ultrafast, with no dates or prices yet.
The bigger point
For a long time the joke about AI design tools was that they could make a beautiful poster for an event that did not exist. Craft gets much closer to the event that does. It understands that a poster has structure, and it usually spells your headline right.
But the reason it understands structure is that it studied finished templates, and finished templates are full of words. So the job has changed shape. You spend less time fighting the layout and more time reading. The designer who wins with Craft is the one who treats every unrequested word as a stranger in the room.
Try it on a real job this week, something small with fixed facts: a menu, a class schedule, an event card. Write down every word that should appear. Generate one round. Then count how many words showed up that were not on your list. That number tells you how much proofreading time to budget, and whether Craft belongs in your comp stage or your finishing stage.