Canva Built Its Own Image Model After Its AI Bill Blew Up

By Daniel Rao

Canva has built its own AI image model — because renting everyone else's nearly broke the business. On Thursday, the Australian design giant launched Craft, the first image-generation model trained from scratch by its in-house lab, Canva Research. It plans a design before it draws it, writes legible text in 55 languages, and has already debuted fourth on the DesignArena leaderboard. The subtext of the launch is the real story: three months ago Canva admitted its AI bill had forced it to slash its 2026 growth forecast from 30% to 20%. Craft is the company's expensive, belated, and entirely necessary answer.

How Canva's Craft plans the picture before it paints it

Most image models turn a prompt straight into pixels. Craft splits the job in two. A planning model first reads the prompt and writes out a detailed blueprint of the composition: which elements appear, where they sit in relation to each other, their relative size, what matters most in the visual hierarchy, and any text that has to appear. A separate rendering model then turns that plan into a finished picture at up to 2K resolution, in 30 aspect ratios or a custom size, as TechCrunch reported from the launch.

Canva says the planning step is what fixes the two things AI image generators are notoriously shoddy at: spelling and layout. Because the structure is decided before rendering begins, text comes out legible — in 55 languages, with prompts accepted in any language — objects land where you asked, and you can edit one element without the rest of the image shifting around. Anyone who has watched a model confidently render a poster headline as gibberish will understand why this matters more than another half-point of photorealism.

Graphic designer Rafael Esquer at work in his New York studio. Craft was trained not on photographs but on millions of human-designed templates — posters, menus, presentations — the raw material of working designers. Photo: RafaelEsquer, CC BY-SA 4.0, via Wikimedia Commons.

Trained on templates, not photographs

Here is the genuinely clever part. Rather than training primarily on photographs, Canva trained Craft on its own library of human-designed templates: posters, social posts, presentations, infographics, menus, typography-rich layouts. Photos teach a model what things look like; templates teach it how layout, hierarchy, and type fit together. That is a structural advantage no general-purpose image lab can easily copy, because nobody else owns a quarter-billion people's design history.

After pre-training, Canva fine-tuned Craft with feedback from its own professional designers — more than 25,000 pairs of images compared against leading models, and, in a detail that tells you everything about the company's priorities, when the automated scores and the designers' opinions disagreed, Canva says it went with the designer. There is something almost old-fashioned about that sentence, and I mean it as a compliment.

The design-first training also feeds Canva's Magic Layers tooling: because Craft understands structure, its flat outputs can be converted into separate, editable design elements inside the editor. The model went live for every Canva user on Thursday inside the Dream Lab tab, and for developers through Canva's public API. The company says it plans to extend generation into third-party chatbots such as ChatGPT and Gemini through its MCP connector — note the verb: plans, not ships — and that two faster, cheaper variants, Craftfast and Craft-ultrafast, are on the way, with no prices or dates attached.

Fourth place, second-best lab — pick your framing

Canva is leaning hard on the independent leaderboard. Craft debuted fourth overall on DesignArena, the crowdsourced benchmark where voters judge images from the same prompts without knowing which model made them. All three models above it belong to OpenAI — GPT-Image-2.5 Sunburst, GPT-Image-2.5 Flare, and GPT-Image-2, as The Verge noted — which lets Canva claim, by its own count, that OpenAI is the only lab ahead of it in image generation. "With Craft, Canva is now the world's second most advanced AI lab for image generation on @DesignArena's global leaderboard," the company posted on X.

The category breakdowns are arguably more impressive than the headline: first out of 87 models in People & Portraits, first out of 85 in Landscapes & Nature, second in Abstract & Patterns. We read Canva's launch post and the leaderboard breakdowns side by side, and the numbers match across the company's own claims and the independent write-ups — which is more than you can say for most launch-day benchmarks. Still, a note of sobriety: fourth is fourth, and "second-best lab" is the kind of framing a marketing department invents when it cannot say "first."

Canva's $10 billion lesson: the AI bill behind Craft

None of this makes sense without the summer Canva just had. In its Q2 2026 investor update, chief executive Melanie Perkins cut the company's 2026 revenue growth forecast from 30% to 20% — and named the culprit plainly. The company had been "relying too heavily on frontier models" from third-party providers, she wrote in her shareholder letter. Several of Canva's own first-party models were not ready when it launched Canva AI 2.0, pricing and usage controls had not caught up with demand, and some AI features reportedly cost six times more to serve than users paid for them.

The numbers were ugly in the way that only private-company numbers can be, because there was no stock price to absorb the shock — only the backers' red pens. Second-quarter revenue came in at $921.9 million, up 25.2% year over year but short of internal expectations, against roughly $4 billion in annualised revenue. The Australian Financial Review reported that Blackbird Ventures and Airtree marked their Canva holdings down by about $10 billion, from $42 billion toward the low-$30 billions. A Nasdaq listing once pencilled for late 2026 slipped into 2027.

Stef Corazza, Canva's head of AI research, has been unusually candid about the economics driving the in-house turn. People make more than a billion designs a month on the platform, he wrote in a post introducing the lab, and "even a variation of a few cents per generation makes an enormous difference." His group has grown past 160 research scientists and engineers inside a 400-person AI unit, absorbing Leonardo.ai — the Sydney startup Canva bought in 2024, whose researchers TechCrunch says contributed to Craft — and the earlier Lucid Origin model was already 5x faster and 30x cheaper than comparable frontier alternatives.

The right humbling

Here is the judgment, and it cuts both ways. Canva's summer was a case study in shoddy cost discipline: shipping AI 2.0 on rented frontier models at unit economics that lost money on every generation is the kind of mistake a company profitable for nine straight years should not make. Perkins called slowing the rollout "a deliberate decision to get the economics right," which is a graceful way of saying they got the economics badly wrong first. The $10 billion markdown was the market's invoice for the lesson.

But the response is the right one, and it is worth saying so plainly. Owning the model is the only way the maths works at a billion generations a month; every rival that keeps renting inference at retail is one pricing change away from Canva's summer. Craft's design-first training is not a gimmick — it is the one moat in image generation that cannot be bought with compute — and the two-stage planning architecture attacks the exact weaknesses, text and layout, that keep AI imagery out of professional workflows. Adobe put OpenAI's models inside Firefly the day they launched; Canva is betting it will not need to.

The open question is whether Craft stays a cost story or becomes a product story. A model that lives inside the editor where a quarter-billion people already work does not need to win the leaderboard to win the market — it needs to make the next poster take five minutes instead of fifty. If Craft does that, nobody will remember it debuted fourth.

Update, Oct. 23: where Craft lives, and what it costs to use

One day in, the practical questions are landing. Craft lives in the Dream Lab tab, which you open from the Canva homepage rather than inside the editor, according to digitaladvisors.com.au's Oct-23 check of Canva's pages — and it is also available to developers through Canva's public API. The Dream Lab help page says some features use your AI allowance and that you can generate up to four images per prompt, with workplace admins able to manage who gets access. What Canva has not said is how many uses Craft takes from that allowance. The price list digitaladvisors saw on Oct 23: Free at $0 gets up to 20 Standard or Premium AI uses a month; Pro at $165 a year for one person gets up to 200 Premium or 20 Ultra; Business at $280 a year per person gets up to 400 Premium or 40 Ultra. The unit-economics question this post opened with — how many cents per generation — stays unanswered on the price side, which is exactly where the $10 billion markdown came from.

Canva Craft: your questions answered

Where do I find Craft in Canva?
In the Dream Lab tab, opened from the Canva homepage (not inside the editor), per digitaladvisors.com.au's Oct-23 check; developers can reach it through Canva's public API.

How many AI credits does each Craft image use?
Canva has not said — its pages state only that some Dream Lab features use the AI allowance and that you can generate up to four images per prompt.

When are Craftfast and Craft-ultrafast coming?
No dates or prices. Canva says only that they will deliver images "at a fraction of the cost of leading frontier models" (via WebProNews).

Sources

TechCrunch (October 22 launch report); Canva's launch announcement (via X); The Verge (via WebProNews); Digital Trends; madrobot; WebProNews; AI Parabellum; Australian Financial Review (via Sprinter/techjuice); Canva Q2 2026 shareholder letter (via StartupDaily/CryptoBriefing). On the economics of AI inference, see also our coverage of DeepSeek's open-sourced inference stack; on AI tooling for storefronts, Shopify letting its AI write your store; and on the agent economy Craft will plug into, Anthropic's Claude Duet and its refusal to touch your money.

Daniel Rao is TechPulse USA's AI correspondent, covering the collision of artificial intelligence with money, power, and regulation. He has strong opinions about unit economics and weak ones about kerning.