Canva shipped Canva Craft on October 22, 2026, the first image-generation model built by its in-house research team, according to TechCrunch. The model targets design work instead of photography. It lives inside Canva’s Dream Lab tab.
Canva says Craft debuted as the fourth most advanced image model on DesignArena, behind three OpenAI models. It outputs images at up to 2K resolution in up to 30 aspect ratios. And it works in two stages, planning a layout before it draws a single pixel.
This analysis covers what Canva announced, how the architecture differs from a standard text-to-image model, where the claims still lack independent proof, and what the launch means for Adobe, OpenAI, Google and the designers who use Canva every day. It also lands during a month of crowded AI launches, which we tracked in our piece on OpenAI Dots versus Meta Muse.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Canva Released on October 22
The announcement is short on pricing and long on positioning. Canva framed Craft as a foundation model trained from scratch, not a fine-tune of someone else’s weights. Thurrott.com quoted the company’s own description of the release, and the wording matters because it makes a specific claim about provenance.
Craft is the first foundation model built and trained from scratch by Canva Research, bringing together years of AI research and investment to deliver a deeper understanding of design, results that more closely match what you imagined, and the freedom to easily edit and refine everything you create.
Canva, company announcement, via Thurrott.com
Two phrases stand out. “Built and trained from scratch” tells rivals and regulators that Canva owns the model end to end. “Edit and refine everything you create” tells customers that the output is meant to be changed, not just admired. Both fit a company whose product is an editor first and a generator second.
| Detail | What was reported | Source status |
|---|---|---|
| Model name | Canva Craft | Confirmed by Canva |
| Builder | Canva Research | Confirmed by Canva |
| Release date | October 22, 2026 | Confirmed by TechCrunch and others |
| Maximum resolution | 2K | Reported by TechCrunch |
| Aspect ratios | Up to 30 | Reported by TechCrunch |
| Where to find it | Dream Lab tab | Reported by TechCrunch |
| DesignArena rank | Fourth at debut | Canva’s claim, not independently verified |
| Price | Not established | Unconfirmed |
How Canva Craft’s Two-Stage Design Works
Most image generators turn a prompt into pixels in one pass. Craft splits the job. In stage one, the model reads the prompt and plans the composition. That plan covers spatial relationships between objects, relative scale, visual hierarchy, and the information needed to render text. In stage two, it generates the image from that plan.
The Verge headlined its coverage around the same idea, saying the model plans designs before generating them. Canva summed up the payoff in a single phrase that The Verge relayed: “Greater precision, consistency and control.”
That phrase reads like marketing, but the engineering logic holds up. Design work fails in specific, repeatable ways. Headlines collide with logos. A menu item drifts off the page. A poster title renders as nonsense letters. A planning step that fixes layout and text placement first attacks those failure modes directly, instead of hoping a single diffusion pass gets everything right.
Why planning first matters for text
Typography is the hardest test for image models. A photo can hide a small error, but a misspelled word on a flyer ruins it. By carrying text-generation information through the planning stage, Craft tries to treat words as layout objects with a size and a position, not as texture. Canva has not published a technical paper in the material we reviewed, so the exact mechanism remains unclear. What the company has stated is the outline of the approach, not the weights or the training recipe.
Built for Design, Not Photography
Canva positions Craft against photography-first models. Reports say it was built specifically for design. That choice shapes everything from training data to the output formats on offer. A model tuned for photorealism optimizes for light, skin and lens behavior. A model tuned for design optimizes for grids, spacing, contrast and readable type.
Training on millions of human-made templates
Canva says it trained Craft on millions of human-made templates from its content library. The company named the categories: posters, social posts, presentations, infographics, menus, and typography-rich layouts. Canva did not publish an exact template count in the reports we reviewed, so “millions” is as precise as the record gets.
This is the strategic core of the launch. Canva sits on a library of professionally structured layouts that rivals cannot scrape from the open web. Each template encodes decisions a human designer made about hierarchy and spacing. A general-purpose image model sees a poster as a picture. A model trained on templates sees it as a composition with rules.
Reports do not say how Canva handled creator consent or licensing for the library, and the fact sheet we worked from does not settle that question. Expect designers and rights groups to ask. We flag it as an open item, not a finding.
2K Output and 30 Aspect Ratios, in Context
TechCrunch reported that Craft generates images at up to 2K resolution and in up to 30 aspect ratios. Those two numbers say a lot about the target market. Canva’s users do not make one kind of image. They make Instagram squares, vertical stories, widescreen slides, A-series print posters and banner ads, often in the same afternoon.
Native support for many aspect ratios saves a step that other tools handle badly. Most generators produce a few standard shapes and then crop or outpaint to reach the rest. Cropping a layout that was planned for a square destroys the composition. Planning for the target shape up front is what a two-stage model can do and a one-pass model struggles with.
The 2K ceiling is a practical limit, not a headline-grabbing one. It covers social media, slides and web banners comfortably. Large-format print would need upscaling. Canva has not said whether Craft will move beyond 2K, and we will not guess.
DesignArena: What a Fourth-Place Debut Tells You
Canva says Craft debuted as the fourth most advanced image-generation model on DesignArena. According to Thurrott.com’s account of the announcement, the three models ahead of it are all from OpenAI: GPT-Image-2.5 Sunburst, GPT-Image-2.5 Flare and GPT Image 2. Canva’s statement also claimed industry-leading visual quality and accurate text rendering.
| Position claim | Model or group | What we can say |
|---|---|---|
| Ahead of Craft | GPT-Image-2.5 Sunburst | OpenAI model, named in Canva’s announcement |
| Ahead of Craft | GPT-Image-2.5 Flare | OpenAI model, named in Canva’s announcement |
| Ahead of Craft | GPT Image 2 | OpenAI model, named in Canva’s announcement |
| Fourth | Canva Craft | Canva’s claim at debut |
| Independent check | None found | Unconfirmed |
Treat the ranking with care. It comes from Canva, it captures a single moment, and leaderboards of this kind shift as voters rate new image pairs. A fourth-place debut is still a strong first result for a company whose core business is not frontier model research. Landing behind three OpenAI models and ahead of the rest of the field would put a first-generation design model in elite company.
The more telling detail is who sits above it. All three leaders belong to a single lab, which tells you how concentrated top-tier image generation has become. Our LLM leaderboard breakdown shows the same pattern in text models, where a handful of labs trade the top spots and everyone else fights for the gaps.
Dream Lab Availability and the Access Question
Craft is available through the Dream Lab tab in Canva, per TechCrunch. The reports split on who can use it. TechCrunch says it is available to all users. Another report says access requires a paid subscription. Canva has not resolved the gap in anything we could verify, so we list exact access and subscription requirements as unconfirmed.
The disagreement may reflect a real structure rather than a reporting error. Many Canva features open to everyone carry usage caps for free accounts, with higher limits behind a subscription. If Craft follows that pattern, both outlets could be describing the same product from different angles. We cannot confirm that, and no price or credit allowance has been published in the coverage we reviewed.
Some coverage, including Tech Times, also mentions developer access through Canva’s API, described as still in preview. Treat that as a report, not a settled fact, until Canva publishes documentation. SiliconReport also covered the launch, and the outlets agree on the core product facts even where they differ on access.
What Is Confirmed and What Is Not
With a launch this fresh, the line between announced and verified matters. The table below separates the two so readers can see which claims rest on named outlets and which still lack proof.
| Claim | Status | Why |
|---|---|---|
| Craft is Canva’s first image-generation model | Reported | Stated by Canva and relayed by TechCrunch and Thurrott.com |
| Built by Canva Research from scratch | Company claim | Canva’s own wording, no outside audit |
| Two-stage plan then generate pipeline | Reported | Described by TechCrunch and The Verge |
| 2K resolution, up to 30 aspect ratios | Reported | TechCrunch |
| Trained on millions of templates | Company claim | No count or licensing detail published |
| Fourth on DesignArena | Unconfirmed | No independent verification found |
| Free versus paid access | Unconfirmed | Outlets disagree |
| Price | Unconfirmed | No figure established |
Why Canva Built Its Own Model
Canva could have kept licensing image models from larger labs. It chose to build one. A Canva spokesperson, quoted by News.com.au, laid out the reasoning: “Building and owning these models gives Canva complete control over cost, quality, and speed, making it possible for us to deploy AI to the hundreds of millions of people who use our products every month.”
Cost, quality and speed are the three levers any platform loses when it rents its core generation capability. Per-image fees from an outside provider scale with usage. A price change or policy shift at the supplier flows straight into Canva’s margins and roadmap. Owning the model turns a variable cost into something the company can optimize with its own engineers.
The same spokesperson said Craft is the first public model from Canva Research, which has tripled in staff numbers in a year, with more models coming soon. That line signals a roadmap, not a one-off. Canva has not named the next model, its modality, or its date, so treat “more coming soon” as intent only.
Competitive Comparison: Own, Rent or Bundle
Craft enters a market where each major player has a different relationship to the model layer. OpenAI sells models and, through its apps, the products on top. Google builds both. Adobe wraps generative tools around its creative suite. Canva now joins the group that owns a model and a distribution channel at the same time.
| Player | Model position | Distribution | Design focus |
|---|---|---|---|
| Canva | Owns Craft, built from scratch | Dream Lab inside the Canva editor | Core purpose, trained on templates |
| OpenAI | Holds the top three DesignArena spots per Canva’s claim | Its own apps and API | General-purpose |
| Builds its own image models | Google products and cloud | General-purpose | |
| Adobe | Firefly generative tools | Creative suite | Professional creative workflows |
| Midjourney | Independent image lab | Its own service | General image generation |
We built this table from public positioning, not from head-to-head testing, so it describes strategy and not quality. For hands-on rankings of the standalone tools, see our roundup of the best AI image generators. The main takeaway is that Canva’s edge is not raw model quality. It is the editor wrapped around the model.
An image from a standalone generator still has to land in a design tool before anyone can use it. Canva generates inside the tool where the text, brand colors and layout already live. That removes an export-and-import cycle for the people who make social posts and slide decks for a living.
Historical Context: From Template Library to Foundation Model
Canva built its business on templates and a simple editor, then added AI features as generative tools matured. Coverage such as Capital Brief’s note on the launch frames Craft as the point where Canva stops depending on partners for its image generation. Earlier AI moves, including its expansion through the Affinity and Leonardo.ai deals, set up that shift. We have not verified acquisition dates or transaction values, so we leave them out.
The pattern has precedent across software. Platforms that start by integrating a supplier’s technology tend to pull the critical piece in-house once usage grows large enough to justify the engineering bill. Cloud providers did it with chips. App makers did it with search and maps. Canva is applying the same logic to generation.
Other platforms are making similar moves with text and code models. Google’s Gemini line keeps expanding inside its own tools, as we covered in our look at Gemini 3.6 Flash in Antigravity. Smaller labs also keep pushing cheaper models, which our report on the MiniMax M3.1 Flash launch explains.
Market Impact for Designers, Marketers and Rivals
For everyday Canva users, the change is practical. A marketer who asks for a menu or a poster gets a draft whose text and layout already respect the design, which cuts the manual cleanup that makes most AI images unusable straight out of the box. That could shift more first-draft work from freelance tools into the editor itself.
For rivals, the pressure lands in two places. Adobe has to prove that Firefly’s integration with professional tools outweighs a cheaper, template-aware model aimed at non-designers. Standalone generators lose some of the casual business audience if the editor they feed into can now generate competitive images natively.
Video is the next obvious front. Creators already stitch image and clip tools together, a workflow we mapped in our guide on how to make an AI video. If Canva Research follows through on its promise of more models, motion could sit on that list, though Canva has not said so.
Risks and Open Questions
Four questions will decide how much this launch matters. First, does the DesignArena result hold once independent voters rate Craft across more prompts? Second, what does access cost, and does the free tier get a meaningful allowance? Third, how did Canva source and license the template library behind the training run? Fourth, can Craft keep text legible across all 30 aspect ratios, not only the common ones?
There is also a quality-versus-control tension. Models that plan layouts can feel rigid, producing tidy but similar results. Designers who want unusual compositions may find a template-trained model drifts toward the familiar. Canva’s emphasis on editing and refining suggests it knows the generated image is a starting point.
Predictions: What Comes Next for Canva and Image AI
- Canva will publish more detail on pricing within weeks. The access confusion is too visible to leave open, and users need to know what Craft costs.
- A second Canva Research model will follow. The company’s own spokesperson promised more models soon. Our bet is on another design-adjacent modality, not a general chatbot.
- Independent benchmarks will narrow the gap or widen it. Third-party testers will rate Craft against the OpenAI leaders. Expect the debut ranking to move in either direction.
- Rivals will stress editor integration in their pitches. Adobe and others will argue that a model alone is not a workflow, and they will tie new generation features closer to editing tools.
- Agent platforms will want Craft access. As assistants from the big labs take on workplace tasks, as in our coverage of Claude Duet versus Muse, a design-grade image model becomes a useful tool for them to call.
These are our forecasts, not reported plans. For more AI model coverage, browse the AI and machine learning section.
Frequently Asked Questions
What is Canva Craft?
Canva Craft is Canva’s first image-generation model, built by Canva Research and released on October 22, 2026. It targets design work such as posters, social posts, presentations, infographics and menus.
Where can I use Canva Craft?
Reports place it in the Dream Lab tab inside Canva. Coverage disagrees on whether every user can open it or only paid subscribers, so check your account when it reaches you.
How much does Canva Craft cost?
No specific price was established in the reports we reviewed. Treat any figure you see as unconfirmed until Canva publishes its own.
What resolution and aspect ratios does it support?
TechCrunch reported output at up to 2K resolution and in up to 30 aspect ratios.
How does Craft differ from other image generators?
It works in two stages. First it plans the layout, including spatial relationships, scale, hierarchy and text information. Then it generates the image. Canva also trained it on millions of human-made templates instead of focusing on photography.
Is Canva Craft the best image model available?
Not by Canva’s own account. The company says Craft debuted fourth on DesignArena, behind three OpenAI models. That ranking has not been independently verified.
Will Canva release more AI models?
A Canva spokesperson said more models are coming soon from Canva Research, though the company has not named them or given dates.
