What Does an Image Model Trained on Canva Templates Actually Make?

Two days after Canva shipped its first image model, reviewer Gaurav Tiwari fed it a café menu brief. Every dish name and every price came back spelled correctly, across all four variations, in under 90 seconds per prompt. One of the poster variations also announced opening hours nobody had asked for. His hands-on review is the most honest summary of Canva Craft so far: it is very good at the part of image generation that usually breaks, and it still makes things up around the edges.

That combination matters more for creators than the leaderboard headlines. Most of us don’t need another photorealism engine. We need thumbnails, promo graphics, event posters and carousel slides where the words are right and the layout holds. Craft was built for exactly that job, and the reason why is also the reason to watch it closely.

A model that learned from posters, not photos

Canva announced Craft on October 22 and made it available to all Canva users globally through the Dream Lab tab, with a public API alongside it. It is the first model built from scratch by Canva Research, a team that grew partly out of the Leonardo.ai acquisition.

The training set is the interesting part. According to TechCrunch, Canva trained Craft on human-designed templates from its own library (posters, social posts, presentations, infographics, menus and typography-heavy layouts) rather than on the photograph-heavy datasets most rivals use. It also works in two stages. A planning pass decides spatial relationships, scale, visual hierarchy and where the text goes; a rendering pass then draws the image at up to 2K.

Plan first, then paint. That is how a designer works, and it is why the text comes out legible. Canva’s developer announcement says Craft accepts prompts in any language and renders in-image text in 55 of them.

What the template diet buys you

Three things stand out for anyone making graphics on a deadline.

Text you don’t have to retype. Headlines, dates and prices landing correctly on the first pass removes the most common reason creators bounce an AI image into a real editor. If you’ve been routing text-heavy graphics through Ideogram for this reason, Craft is now a credible alternative that lives where you already design.

Output that becomes an editable design. Images from Craft can be converted into layered designs with Magic Layers, so the headline becomes a text box you can fix and the graphics become separate elements. Tiwari’s menu converted in about a minute (the dotted leader lines didn’t survive). That is the same direction we saw with editable image layers in Seedream and Reve, and with Adobe’s editing controls inside ChatGPT: the industry is admitting a flat PNG is the wrong deliverable for design work.

A real ranking, with an asterisk. Craft debuted fourth on the DesignArena image leaderboard. The three models above it, per Tech Insider, are all OpenAI: GPT Image 2.5 Sunburst, GPT Image 2.5 Flare and GPT Image 2. That is what lets Canva call itself the second-best image lab. The asterisk: Craft had about 2,082 head-to-head matches at launch against roughly 61,500 for GPT Image 2, and its score slipped from 1,340 to 1,330 in a day, per this leaderboard breakdown. Treat fourth place as a first impression, not a verdict. (If you’re weighing the OpenAI side, our GPT Image 2.5 Flare vs Sunburst breakdown covers when each one earns its cost.)

Where it still slips

The same hands-on test turned up five problems worth knowing before you build a workflow on Craft.

1. It invents copy. The coffee poster gained an unrequested opening-hours footer in one variation, plus taglines and badges in others. Proofread every word, including the ones you didn’t write.
2. Your Brand Kit doesn’t apply yet. The “Apply brand” option was grayed out in Dream Lab during testing. For creators whose whole point in Canva is brand consistency, that is a big gap.
3. Fewer ratios than advertised. Canva promotes 30+ aspect ratios, but Dream Lab showed six in Tiwari’s test, and the 9:16 output wasn’t an exact 1080 × 1920 story size. The API preview is limited to square, landscape and portrait for now.
4. Small text stays small. Infographic step descriptions were too tiny to read on a phone. A model trained on print-style layouts doesn’t automatically know your audience is scrolling.
5. Edits re-render the whole image. Prompt edits caused faint pixel differences across untouched elements. This is the same drift problem we covered with Krea 3; one change per edit and a close comparison afterward is still the safe habit.

There’s a quieter risk too. A model trained on Canva templates has learned what Canva templates look like, and 250 million monthly users have been publishing those templates for years. If your thumbnail is supposed to stop the scroll, “looks like a polished Canva design” may be the problem, not the goal. Use Craft for structure, then push the style somewhere the template library hasn’t been.

The meter Canva hasn’t printed

Canva has not published what a Craft image costs. It draws from the shared AI allowance on your plan: up to 20 Standard or Premium AI uses a month on Free, up to 200 Premium uses on Pro and up to 400 on Business, according to Canva’s help page as summarized in this launch analysis. Which tier Craft falls into, and how many uses one generation counts as, isn’t stated. Discarded variations still count.

The motive behind the model explains why that number matters. In August, Canva cut its revenue growth forecast from 30% to 20% and co-founder Mel Perkins said the company had been “relying too heavily on frontier models” because serving AI tasks cost too much. Craft is, in part, Canva taking its image bill in-house, and the planned Craft-fast and Craft-ultrafast variants are pitched as cheaper still. That is good news if the savings flow into bigger allowances. It is a reason to watch your usage page if they don’t.

Two smaller rights notes. Canva permits commercial use of AI output, but Tiwari notes its AI indemnity covers only eligible Enterprise customers, not Pro or Business, and purely AI-generated images may not qualify for copyright at all (our AI image copyright explainer covers why). And if you use the API, download links expire after 24 hours, so save outputs to your uploads.

Which jobs to hand it this week

• Hand it: event posters, menus, quote cards, carousel slides, simple infographics and promo graphics where exact words matter and you’ll finish the layout in Canva anyway.
• Test it against your current tool: YouTube thumbnails. Pair it with a face shot you took yourself, and compare against YouTube’s own AI thumbnail tools before you commit.
• Keep elsewhere for now: anything that must match a strict brand kit, story-size assets that need exact pixel dimensions, and work you plan to iterate on heavily through prompt edits.

On the Free plan, 20 monthly uses is enough for one honest test. Pick the graphic you remake most often, run the same brief through Craft and through whatever you use today, and count how many minutes of cleanup each one needs. That number tells you more than the leaderboard does.