Canva Craft: Canva's own image model born already connected to Magic
On October 22, 2026, Canva launched its first in-house image generation model, Canva Craft, designed to deliver not a standalone image but a design that's already editable in layers within the tool itself.
On October 22, 2026, Canva launched its first in-house image generation model, Canva Craft, designed to deliver not a standalone image but a design that's already editable in layers within the tool itself.
Editor's note: this article cites financial projections (Canva's valuation dropping from 42 billion to 35 billion dollars) attributed to "external analysts" without named identification in the original report, and there is no public benchmark comparing Craft to competing models. We kept this information because it does not contradict the core of the news, but readers should treat it as market context we cannot audit, not as verified technical data.
On October 22, 2026, Canva launched its first in-house image generation model, Canva Craft, designed to deliver not a standalone image but a design that's already editable in layers within the tool itself.
On October 22, 2026, Canva launched its first in-house image generation model, Canva Craft, designed to deliver not a standalone image but a design that's already editable in layers within the tool itself.
Canva announced Canva Craft on October 22, 2026, its first image generation model developed in-house by the company's research team. The news, published by TechCrunch, positions Craft as a direct competitor to the image models from Google, OpenAI, ByteDance, and Meta. But the detail that matters to those designing products is not the competition between models: it's where the generated image goes once it's ready.
What each tool does on its own, and what the combination unlocks
On its own, Craft is just another prompt-based image generator, albeit with a particular architecture. It works in two stages: first it interprets the user's request and builds a plan with the spatial relationships between elements, relative scale, visual hierarchy, and the text instructions that will appear in the piece; then, based on that plan, it renders the final image, with resolution of up to 2K and 30 different aspect ratios.
Canva Magic, in turn, doesn't generate anything: it interprets a ready-made image and transforms it into a design that's editable in layers. Canva describes Magic as an image-to-design transformer that "understands designs and layers," without publicly detailing how this specific separation (text, background, graphic elements) happens internally. Magic launched in 2025.
The combination of the two is the central point of the launch: according to Canva, images generated by Craft can be converted into layered designs using Magic. In practice, this changes the type of result a generative AI delivers within a design workflow: instead of a flattened PNG that needs to be redesigned from scratch to become editable, the designer gets a starting point that's already structured into manipulable elements.
Prerequisites: plans, access, and cost
Craft is available to all Canva users, accessible through the platform's Dream Lab tab, with no indication of restriction to paid plans in the source consulted. It's also available via the Canva API, a relevant path for those building products on top of the platform and not just designing within it. Canva also plans to support image generation in third-party chatbots, such as ChatGPT and Gemini, through Canva MCP, but this integration is described as a future plan, not a feature already available.
On direct cost to the end user, the source doesn't detail figures. What it confirms is the business rationale behind the decision to build its own model: Canva already used several third-party image models on its platform, and Craft is expected to help reduce API costs, in addition to giving the company more control over generation quality and speed.
This control matters especially because Canva revised its 2026 growth projection downward, a move that, according to external analysts cited by TechCrunch, dropped the company's valuation from 42 billion dollars to 35 billion dollars, even with revenue projected at 4 billion dollars the previous year.
In short: bringing the image model in-house is, above all, a margin decision at a time of slower growth, not just a product quality decision.
Step by step: from prompt to editable design
Imagine a marketing designer who needs to put together an Instagram carousel for a clearance campaign, with a square-format piece and legible highlighted text. The workflow Canva describes for this scenario follows this sequence:
- Open Canva and access the Dream Lab tab, where Craft is available to all users.
- Write the prompt describing the piece: for example, a clearance poster with a purple background, a highlighted "50% OFF" headline, and a minimalist style, and choose the ratio among the 30 supported aspect ratio options.
- Craft runs the first stage, planning where the text goes, which element has the most visual hierarchy, and what scale each object occupies in the composition.
- In the second stage, the model renders the final image at up to 2K resolution, already with the "50% OFF" text legible and positioned according to the plan.
- This generated image can then be converted by Canva Magic into a layered design, since Magic is described by Canva as a transformer that understands designs and layers, although the company doesn't publicly detail how this specific separation happens.
- From there, the designer edits normally in the Canva editor: changes fonts, adjusts colors, resizes text, without needing to generate the entire image again.
An architecture detail explains why this workflow works better for design pieces than for photographic images: Canva trained Craft using human-designed templates from the platform's own library, such as posters, social media posts, presentations, infographics, menus, and text-heavy layouts, instead of using photographs as training data. Researchers from Leonardo.ai, a media generation company that Canva acquired in 2024, contributed to the model's development.
Where it breaks down
The first limitation is availability: Canva confirms it's developing two cheaper versions of the model, called Craftfast and Craft-ultrafast, but only says they "will be released soon," with no date. This means that, today, anyone who needs fast, cheap generation at volume still doesn't have this option available, and Craft's real cost-benefit at scale remains publicly untested.
The second is about external integration: support for image generation inside chatbots like ChatGPT and Gemini via Canva MCP is an announced plan, not a feature in production. Teams that were hoping to generate Canva pieces from within other AI tools still can't do that.
A third, which TechCrunch doesn't address directly but which follows from the training decision itself, is the model's reach outside the graphic design universe: a model trained on human templates tends to be stronger at legible text, hierarchy, and piece composition than at scene photorealism, an area where models trained primarily on photography usually have the advantage. The source doesn't publish any benchmark comparing Craft's quality to that of the models from Google, OpenAI, ByteDance, or Meta, so any claim of technical superiority, in either direction, is still a launch promise, not a measured result.
| Model | Current stage | Stated focus |
|---|---|---|
| Canva Craft | Available in the Dream Lab tab and via API | Quality and legible text in design pieces |
| Craftfast | In development, no date | Reduced-cost version ("cost-efficient"), with no focus differentiation detailed by the source |
| Craft-ultrafast | In development, no date | Reduced-cost version ("cost-efficient"), with no focus differentiation detailed by the source |
Who it's for and who it's not for
The Craft plus Magic combination is worthwhile for those who already live within the Canva ecosystem and put together template-heavy pieces day to day, such as marketing and product designers and product ops teams who need posts, presentations, and campaign materials ready to edit without going through another tool. It's also worthwhile for those who build products on top of the Canva API and want to generate programmatic design without maintaining their own integration with third-party models.
It's not worthwhile, at least for now, for those who need high-fidelity photorealistic generation, since the model was trained on templates rather than photographs, nor for those who depend on low cost at high volume, since the cheaper versions still haven't left the drawing board. It also doesn't solve, today, the use case of generating pieces from inside ChatGPT or Gemini, because support via MCP remains a roadmap item, not a product. For those deciding where to invest team time now, the safer bet is to test the Craft plus Magic workflow on internal pieces before relying on it for critical client deliverables.
Source
Ivan Mehta, "Canva releases its first image generation model," TechCrunch, October 22, 2026. Available at: https://techcrunch.com/2026/10/22/canva-releases-its-first-image-generation-model/.
Translated from the Brazilian Portuguese original · Read the original
Stripe and OpenAI define who controls checkout when the AI agent buys for you
The Agentic Commerce Protocol, published by Stripe and OpenAI in September 2025, doesn't just solve the technical part of the agent that buys: it allocates, by contract, who is the merchant of record, who holds the customer's data, and who processes the payment.
