Canva released Canva Craft, its first image-generation model developed by its research team, on 22 October 2026, TechCrunch reported. The design platform has made the model available to all users through its Dream Lab tab, bringing an image generator built by Canva into a product that has previously used models from other companies.
Key points
- Canva says Craft plans the arrangement and scale of elements before generating an image of up to 2K resolution.
- Preliminary blind comparisons placed Craft fourth on Design Arena’s image-generation leaderboard.
- All Canva users can access Craft in Dream Lab. Access through the Canva API and third-party chatbots is planned.
- Canva trained the model on human-designed templates from its library, including posters and presentations.
Canva Craft plans the layout first
Canva describes a two-stage process. Craft first reads a prompt and works out where the requested elements should sit, how large they should be relative to one another, which should draw attention and what text the image needs. It then generates the image, at up to 2K resolution in a choice of 30 aspect ratios. The plan is akin to marking out the space for words and pictures on a poster before filling in their details.
Making a poster could start with a description of how its text and other elements should be arranged. If the text came out legible and an edit stayed within the part being changed, the rest of the layout could remain in place. Those are capabilities Canva claims for Craft, rather than outcomes guaranteed for every image.
The company says the model can produce readable text and accurate object details, and can make a precise change without altering other parts of an image. Those claims matter especially for layouts that have to carry information: a word in a menu or a presentation needs to work as text, while the surrounding design may need to survive an adjustment.
Canva says it trained Craft on human-designed templates in its own library, including posters, social posts, presentations, infographics, menus and layouts rich in typography, rather than using photographs as its training material. That choice fits the tasks the company describes for the model: arranging elements and placing words within an image, as well as generating its visual content.
TechCrunch reported that Craft images could be converted into layered designs with Canva Magic, Canva’s image-to-design transformer. Layers matter because they keep parts of a composition available for separate changes, instead of treating every requested edit as a change to a single finished picture. Canva released a model that understands designs and layers last year.
Design Arena places Craft fourth
Preliminary results placed Craft fourth on Design Arena’s image-generation leaderboard, ahead of models from Google, Microsoft, Meta, xAI and ByteDance, Capital Brief reported. The independent benchmark uses blind comparisons: images produced from identical prompts are judged without identifying the model behind each one.
Capital Brief reported that Canva brought its announcement forward after those preliminary results. It also placed Canva second among AI laboratories for image generation, behind OpenAI, when the companies were ranked by their models’ performance on that benchmark. The fourth-place result ranks models; the second-place result ranks the organisations behind them.
The blind comparisons concern the images generated from the same prompts. Canva’s account of Craft’s internal planning, its training material and its editing abilities comes from the company. For a developer choosing an image model, those are different pieces of information: one records preferences in a particular comparison, while the others describe how Canva says the model can be used in design work.
Dream Lab opens Craft to Canva users
Craft is available to all users in Dream Lab on Canva’s platform. Canva says it will also make the model available through its API, which would let software built on Canva call the generator directly. The company plans to support image generation in third-party chatbots, including ChatGPT and Gemini, through Canva’s MCP.
Canva has used several other image-generation models on its platform. The company says owning Craft gives it more control over quality and speed. TechCrunch reported that using its own model would probably also reduce the API costs associated with relying on outside models. Canva says it is developing two cost-efficient variants, Craftfast and Craft-ultrafast.
The release comes as Canva’s growth outlook has changed. The company had been on track for $4 billion in revenue last year, but cut its growth forecast for 2026. TechCrunch reported that, according to external analysts, its valuation then fell from $42 billion to $35 billion.
Canva acquired Leonardo.ai, a media-generation platform, in 2024. Canva says researchers from Leonardo.ai contributed to Craft.
