Canva spent about six months building its first image model from nothing. That is the detail Australian marketers and design teams should hold onto this week. Canva Craft went public on 22 October 2026, and Australian coverage, including our launch report on its DesignArena ranking, followed on 23 October. This piece looks at what the build tells us about Canva’s strategy.
Canva calls Craft its first public image-generation model and its first foundation model developed from scratch. Canva Research, the company’s in-house AI team, built it. Co-founder Cliff Obrecht put the team at 160 people. The model landed at fourth place on DesignArena, a human-evaluated benchmarking platform for AI design. Pricing is still unconfirmed, and so is the broader claim that Craft beats larger rivals.
The launch matters beyond one product. A design platform with a huge template library has decided to train its own model on that library instead of renting someone else’s. Startup Daily was among the outlets that reported the home-grown model on launch. This analysis separates what Canva has confirmed from what outlets have inferred, then works through the market effects.
Don’t miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
Add Now
What Canva Launched on 22 October 2026
Craft is a text-to-image model. Users reach it through Dream Lab, Canva’s generation space for its own customers. Developers and AI agents reach it through Canva’s public API. That dual route is the first signal of intent. Canva is not treating Craft as a hidden feature inside the editor. It is treating it as a product other software can call.
Output tops out at 2K resolution, in more than 30 aspect ratios. Canva also says the model renders text in 55 languages. Text rendering is the weak spot of many image generators, so a language claim of that scale is a deliberate pitch to designers who ship posters, social posts and slides rather than artwork.
Two follow-up models are in development, called Craft-fast and Craft-ultrafast. Canva has not published release dates or pricing for either. The table below lists what the available coverage confirms and how firm each item is.
| Item | What was reported | Status |
|---|---|---|
| Model name | Canva Craft | Confirmed |
| Developer | Canva Research, in-house AI team | Confirmed |
| Launch date | 22 October 2026 (Australian coverage 23 October) | Confirmed |
| Access | Dream Lab and Canva’s public API | Confirmed |
| Maximum output | 2K resolution, 30-plus aspect ratios | Confirmed |
| Text rendering | 55 languages | Canva claim |
| DesignArena rank | Fourth | Reported |
| Price | Not published | Unconfirmed |
Six Months, 160 People: Reading the Build Speed
Obrecht described Craft in direct terms. In his words, “Craft is the first foundation image model from our 160-person Canva Research team, built completely from scratch in about six months.” The phrase “from scratch” carries weight. It tells readers Canva did not fine-tune an open model or wrap a licensed one.
Six months is short for a foundation model. Canva has not said how much of that time went to data preparation versus training, and it has not published compute figures. Treat any outside estimate of cost as speculation. What the timeline does show is that Canva had a clear dataset advantage and a team ready to use it.
Canva’s own statement on growth adds context. The company said Craft is the first public model from Canva Research, “which has tripled (staff numbers) in a year, with more models coming soon.” A team that tripled in twelve months did not start small. Reports differ slightly on size, with one describing more than 160 scientists and engineers, so the conservative figure is the 160 Obrecht gave.
Why Canva Trained on Templates Instead of Photographs
Canva says Craft learned from millions of human-made design templates in its content library. Those include posters, social posts, infographics and presentations. Canva says the model did not train primarily on photographs. That choice separates Craft from generators built on scraped image-and-caption pairs, and it shapes what the model is good at.
Templates carry information that photographs lack. They encode hierarchy, spacing, type pairing and the way a headline sits against a block of body copy. A model that sees millions of finished layouts can, in principle, learn what a usable poster looks like and not only what a pretty picture looks like. Canva has not published evaluation data proving this, so it remains a design hypothesis the benchmark partly tests.
The data-rights question
Training on a company’s own content library raises a rights question. Many templates come from contributors and designers, and the available reports do not explain how Canva’s licensing terms apply to model training. That gap matters in a country where creators are already pressing for rules. Australian authors have pushed for a dedicated law, as we covered in our report on the 98% author push for AI copyright reform.
Music rights holders have raised similar alarms. Music publishers’ body ICMP warned about large training datasets, which we detailed in our piece on Kylie Minogue songs in a 12M AI dataset. Canva’s approach, training on content it holds in its own library, sits closer to a controlled dataset. Whether contributors agree is a separate matter that Canva has yet to address in the coverage reviewed here.
The Two-Stage Design: Planning Separate From Rendering
Craft uses a two-stage system. One stage handles compositional planning, meaning where elements go and how the layout holds together. The other stage renders the image. Canva has not published a technical paper in the sources reviewed, so the internal details stay unconfirmed.
The logic is easy to follow. Layout is a structural problem, and pixels are a rendering problem. Splitting them lets each stage specialise. It also fits Canva’s core business, which has always been about arranging text, shapes and images on a page. An image model that plans the page first mirrors how a human designer works.
For buyers, the practical question is control. A planning stage could, in theory, expose layout to users or agents before rendering. Canva has not said that it does. Until documentation appears, developers should assume the API returns finished images and test whether they can steer composition through prompts alone.
Where Craft Landed on DesignArena
DesignArena is a human-evaluated platform, so rankings reflect people choosing between outputs instead of automated scores. Craft placed fourth. That is a strong debut for a first public model, and it is also not first place. Coverage that says Craft “outperforms” Google, Microsoft, xAI, Meta or ByteDance goes further than the evidence supports.
A fourth-place finish tells us three things. Canva reached the top tier of design-focused generation on its first attempt. At least three models rank above it on this benchmark. And the ranking comes from one platform, so it says little about photorealism, video or other tasks Craft was not built for.
Readers choosing an image tool should weigh that carefully. Benchmarks that focus on design tasks favour a model trained on design data. Our comparison of Nano Banana Pro, Photoshop AI and GPT-Image shows how much results shift once the task changes from layout to photographic detail.
55 Languages and 30-Plus Aspect Ratios in Practice
Text inside images has been a persistent failure point for generators. Letters warp, words misspell and non-Latin scripts fare worst. Canva says Craft handles text in 55 languages. That claim comes from Canva and has not been independently tested in the sources reviewed, so teams should run their own checks with the languages they actually use.
Aspect ratio support matters for the same reason. A single campaign may need a square social tile, a vertical story, a widescreen banner and a slide. More than 30 ratios means fewer crops and fewer awkward regenerations. For agencies that resize one concept across a dozen placements, that cuts steps rather than adding features.
The 2K ceiling is a limit worth noting. Print work and large-format signage may need more pixels, and creators would have to upscale elsewhere. For screens, social feeds and presentations, 2K covers most needs.
Dream Lab, the API and the Agent Angle
The API release may prove more consequential than the consumer feature. Canva says developers and AI agents can call Craft directly. As automated workflows grow, agents that draft a campaign, write copy and generate artwork will need an image endpoint tuned for layouts. Canva is positioning Craft for that job.
Secondary reports say Craft is available to every Canva user. Canva’s materials do not spell out plan eligibility, usage limits or fair-use caps, so that claim is only partly confirmed. Free-tier users should expect limits, though the size of those limits is unknown.
What developers should test first
- Text accuracy in the specific languages your product serves
- Output consistency across repeated prompts with the same brief
- Latency, since Craft-fast and Craft-ultrafast are not yet released
- Rate limits and pricing once Canva publishes them
How Craft Compares With Rival Image Models
A fair comparison has to stay inside confirmed facts. Canva has not published head-to-head numbers against named rivals, and DesignArena gives rank without a score gap in the reports reviewed. The table frames the strategic differences instead of inventing performance figures.
| Factor | Canva Craft | Typical general-purpose generator |
|---|---|---|
| Training emphasis | Millions of design templates | Broad image data, often photo-heavy |
| Core strength pitched | Layouts, text, posters, slides | Illustration and photorealism |
| Text languages claimed | 55 | Varies by model |
| Delivery | Dream Lab and public API | Own app, API or both |
| DesignArena rank | Fourth | Varies |
| Pricing | Unconfirmed | Published per model |
Adobe’s Firefly, Midjourney and Leonardo AI remain the names most Australian creatives already know. Our breakdown of Midjourney, Leonardo AI and Firefly pricing shows how wide the cost spread is, which sets the bar Canva must clear once it publishes Craft pricing. Open models are another threat. Teams comfortable with local tooling can follow our Qwen-Image-2.1 ComfyUI setup guide and run generation at zero licence cost.
Market Impact for Australian Designers and Marketers
Canva is an Australian company, and its model launch lands in a market where small businesses, schools and agencies already live inside its editor. Adding a native generator removes a step for them. A marketer no longer needs a separate subscription to create a hero graphic, then import it, then lay out text.
The bundling effect is the main commercial point. Standalone generators compete on image quality alone. Canva competes on image quality plus layout, brand kits, collaboration and publishing in one place. A model that scores fourth on a design benchmark but sits inside the tool people already use can win on convenience. That dynamic helps explain why several platform vendors have chosen to build rather than license.
Freelance designers face a mixed picture. Routine work such as social tiles and simple flyers gets cheaper to produce. Higher-value work such as brand systems and art direction depends on judgement a model does not supply. Agencies will likely use Craft to speed first drafts, then refine by hand.
For wider coverage of model launches and benchmark shifts, see the AI and machine learning section on Tech Insider.
Historical Context: From Template Library to Model Builder
Canva built its audience on templates and a drag-and-drop editor, and its content library became the raw material for Craft. Canva describes Craft as its first foundation model and its first public image model. That is a change in posture, from a design tool that offers AI features to a company that trains its own models.
The pattern is familiar across the sector. Platforms with proprietary data often reach a point where licensing a general model costs more in control than it saves in effort. Owning the model lets Canva tune behaviour for its own content, set its own pricing and avoid dependence on a supplier that could change terms. Canva has not said this was the reason, so treat it as industry logic and not a stated motive.
The staffing history supports the reading. Canva says its research team tripled in a year. Hiring at that pace points to a plan that predates this launch, and the “more models coming soon” line suggests Craft is a first release rather than a one-off.
What Remains Unconfirmed
Several headline claims need caution. The table separates what Canva has put on the record from what commentators have added.
| Claim | Evidence available | Our view |
|---|---|---|
| Craft beats Google, Microsoft, xAI, Meta, ByteDance | Fourth place on one benchmark | Not established |
| Available to every Canva user | Secondary reports only | Partly confirmed |
| Craft-fast release date | None published | Unconfirmed |
| Craft-ultrafast release date | None published | Unconfirmed |
| Pricing for API and plans | None published | Unconfirmed |
| Template contributor consent for training | Not addressed in coverage reviewed | Open question |
Readers should hold off on budget decisions until Canva publishes API pricing and plan limits. Early adopters can still test Craft in Dream Lab to judge quality against their own briefs.
Five Predictions for Craft and Canva Research
These are analyst-style forecasts based on the confirmed facts, not reported plans. Canva has not confirmed any of them.
- Craft-fast ships first. Canva named the lighter models, so expect at least one to reach users within months, aimed at speed-sensitive features inside the editor.
- API pricing undercuts general generators. A model trained on owned data and delivered at scale gives Canva room on price, and it will want developers building agents on Craft.
- Benchmark ranking becomes a marketing line. Canva will push the fourth-place result and aim to climb, so watch DesignArena for movement after the next release.
- Rights scrutiny follows. Template contributors and creator groups will ask how training data was licensed, given the local debate over AI and copyright.
- Rivals respond with layout features. Competing image tools will add stronger text and layout controls to protect design-focused users.
What to Watch Next
Four signals will show whether Craft is a lasting product or a launch-week story. First, Canva’s pricing page for the API. Second, any technical write-up of the two-stage architecture. Third, the next DesignArena standings after Craft-fast arrives. Fourth, any statement on contributor consent for template training.
The takeaway: Canva proved it can build a competitive image model quickly, but proof of leadership is still missing. A fourth-place debut from a six-month build is a serious opening move. The next release will show whether it is the start of a climb.
Frequently Asked Questions
What is Canva Craft?
Canva Craft is Canva’s first public image-generation model and its first foundation model built from scratch. Canva Research developed it, and it launched on 22 October 2026.
How long did Canva take to build Craft?
Co-founder Cliff Obrecht said the model was built completely from scratch in about six months by a 160-person research team.
Where can I use Craft?
Craft is available through Canva’s Dream Lab for users and through Canva’s public API for developers and AI agents. Plan eligibility and usage limits are not clearly specified.
What resolution and formats does Craft support?
It produces images up to 2K resolution in more than 30 aspect ratios. Canva also says it can render text in 55 languages.
What data was Craft trained on?
Canva says it trained Craft on millions of human-made design templates from its content library, including posters, social posts, infographics and presentations, rather than primarily on photographs.
Is Craft better than Google, Microsoft or Meta image models?
Not proven. Craft ranked fourth on DesignArena, a human-evaluated benchmark. Claims that it outperforms named rivals go beyond the evidence in the available reports.
How much does Craft cost?
No confirmed price has been published. Pricing for the API and for the planned Craft-fast and Craft-ultrafast models is also unconfirmed.
Are more Craft models coming?
Canva has said it is developing Craft-fast and Craft-ultrafast, and says more models are coming soon. Release dates have not been announced.
