Canva Just Built Its Own AI Image Model — And It Already Beats Google and Meta

On Thursday, October 22, 2026, Canva — the design platform used by more than 250 million people — announced something it has never done before: it built its own AI image generation model from scratch.

It's called Canva Craft, and it's already making waves. In early blind comparisons on Design Arena's public leaderboard, Craft debuted at 4th place — ahead of image models from Google, Microsoft, Meta, xAI, and ByteDance, trailing only OpenAI. For a company's first-ever foundation model, that's a remarkable entrance.

Key Takeaways

• Canva Craft is Canva's first in-house AI image model, built from scratch by Canva Research (with help from the Leonardo.ai team Canva acquired in 2024).
• It works in two stages: first it plans the layout (composition, text, spacing), then it renders the image at up to 2K resolution across 30+ aspect ratios.
• Trained on human-designed templates — not photographs — so outputs convert into editable, layered Canva designs instead of flat pictures.
• It renders legible text in 55 languages, a notorious weak spot for most AI image generators.
• Available now to all users in Dream Lab and via the Canva API; faster variants (Craftfast, Craft-ultrafast) and ChatGPT/Gemini integrations are coming.[

Image: Canva logo (via thebytenerd.com)

Why Canva stopped renting AI models

Until now, Canva has powered its AI image features with other companies' models. When Dream Lab launched in October 2024, it ran on Phoenix — the model from Leonardo.ai, the Sydney startup Canva acquired that same year. Every image generated meant paying someone else for the compute.

That arrangement was getting expensive. Canva reportedly cut its 2026 revenue growth target from 30% to 20%, with AI infrastructure costs cited as a factor. Building Craft in-house is a direct answer: own the model, own the costs, own the roadmap. As one analyst put it, Canva just stopped renting and bought the house.

The Leonardo.ai researchers didn't leave, either — Canva confirmed to TechCrunch that the Leonardo team contributed directly to Craft's development. This is an acquisition paying off in the most literal way: the team Canva bought two years ago just built its flagship AI model.

How Craft actually works: plan first, paint second

Most AI image generators work in one shot: you type a prompt, the model turns it directly into pixels. Craft splits the job into two distinct stages, and that architectural choice is the whole story.

Stage one: the planner. A planning model reads your prompt and writes out a detailed composition plan — which elements appear, where they sit relative to each other, their relative sizes, what matters most visually, and any text that needs to appear in the image. Think of it as an art director sketching the layout before anyone picks up a brush.

Stage two: the renderer. A separate rendering model takes that plan and produces the final image at up to 2K resolution, in more than 30 aspect ratios or a fully custom size.

Why does this matter? Three reasons. First, text actually works — Craft can write legible text into images in 55 languages, solving one of the most embarrassing failure modes of AI image generators (garbled, misspelled, or alien text). Second, objects land where you asked — the planning stage enforces spatial relationships instead of hoping the model figures them out. Third, you can edit one element without redrawing everything — because the image has an underlying structure, not just pixels.

Image: Canva Dream Lab (via kawebsites.gr)

Trained on designs, not photos — and that's the point

Here's the detail that separates Craft from every other major image model: Canva trained it on human-designed templates from its own library — posters, infographics, social media posts, presentations, menus, typography-heavy layouts — instead of the photographs and web-scraped images that models like DALL-E and Midjourney learn from.

This is a deliberate bet on Canva's home turf. A model trained on photographs is great at making things look real. A model trained on designs is great at making things look designed — understanding visual hierarchy, white space, typography, and brand consistency from the start.

And it unlocks Canva's killer feature: because Craft's outputs carry design structure, they feed directly into Magic Layers, Canva's tool that converts a flat image into separate, editable design elements. Most AI generators hand you a finished picture — take it or leave it. Craft hands you a starting point you can actually work with inside Canva's editor. For Canva's 250+ million users, most of whom are not professional designers, that difference is everything.

4th place on debut: what the ranking really means

Canva claims Craft is the world's second-best AI lab for images behind OpenAI, citing Design Arena's public leaderboard — though the preliminary blind-comparison data reviewed by independent outlets places it 4th overall, ahead of Google, Microsoft, Meta, xAI, and ByteDance.

A grain of salt is warranted: the ranking comes from preliminary blind comparisons, not a finalized leaderboard position, and Canva has not published a technical report with parameter counts or training details. But even a 4th-place debut for a first-ever model — from a design company, not an AI lab — signals that the "plan first" architecture is genuinely competitive.

The honest caveat from early testers: Craft's photorealism still lags behind DALL-E 3 and Midjourney on complex scenes. It excels at graphic design, marketing materials, and social assets — exactly what it was built for — but don't expect it to win a photography contest yet.

What's next: faster, cheaper, everywhere

Canva isn't stopping at one model. Two variants are already announced: Craftfast (a faster, cheaper version) and Craft-ultrafast (even more lightweight). The names tell you the strategy — Canva wants Craft running everywhere, at every price point, inside every workflow.

Distribution is equally aggressive. Craft is live now for all users in the Dream Lab tab and through Canva's public API for developers. Next on the roadmap: image generation inside third-party chatbots including ChatGPT and Gemini via Canva's MCP server. Canva isn't just building a model — it's building a model platform.[

Image: Canva design workspace (via website-files.com)

What this means for you

You don't need to be a Canva user to feel this shift. Here's the practical impact:

• Better AI images in the tools you already use. If you use Canva (and statistically, you probably do), the image generator just got smarter, faster, and more editable — at no extra cost.
• Text in AI images finally works. Posters, social graphics, and thumbnails with readable headlines — in 55 languages — without the garbled-text lottery.
• The "design AI" category is now real. Until Craft, every major image model was a generalist trained on photos. Canva just proved there's a market for a specialist trained on design — expect competitors to follow.
• Cheaper AI for everyone, eventually. When a company with 250 million users stops paying third-party API fees, those savings eventually flow into the product — better free tiers, more features, lower prices.

The takeaway

Canva Craft is more than a product launch — it's a statement of intent. A design company just built a frontier AI image model from scratch, debuted at #4 on the leaderboard, and made it free for a quarter-billion users. The two-stage "plan first, paint second" architecture and the template-trained approach give it a genuine edge in the design workflows where most people actually use AI images.

The AI image race just got a new runner — and it came from the design world, not the AI lab world. That should worry everyone else.

Read more: detailed breakdown of the Canva Craft launch and SiliconReport's coverage.