Canva Craft: a two-stage image model that plans a design before rendering

Canva released Canva Craft, an in-house image model that builds a layout plan first, then renders up to 2K across more than 30 aspect ratios.

What happened

Canva has launched Canva Craft, its first in-house image generation model. The Verge reports the model competes directly with those from OpenAI, Microsoft, and Google; TechCrunch lists Google, OpenAI, ByteDance, and Meta as same-category competitors.

Canva says the model works in two stages. According to TechCrunch, the first stage interprets the user's prompt and creates a plan covering spatial relationships of elements, relative scale, visual hierarchy, and information for text generation. The second stage generates an image at up to 2K resolution in a possible 30 different aspect ratios. Canva says this gives "greater precision, consistency and control" and helps with complex images that require legible text and precise object placement, and that users can make precise edits without changing other parts of the image.

Canva says it trained Craft on "millions of human-made templates" from its content library, including posters, social posts, presentations, infographics, menus, and typography-rich layouts, rather than on photographs. TechCrunch notes images generated with Craft can be converted into layered designs using Canva Magic, the company's image-to-design transformer.

The Verge reports Craft is currently ranked fourth on Design Arena's crowdsourced leaderboard for AI image generation models, behind OpenAI's GPT-Image-2.5 Sunburst, GPT-Image-2.5 Flare, and GPT-Image-2. Canva says the model ships with "robust input and output moderation, privacy and safety controls, and reporting tools."

Craft is available now to all users globally through the Dream Lab tab, reachable via the three-dot option on Canva's left-hand menu, and through the Canva API. Canva acquired Leonardo.ai in 2024 and says Leonardo researchers contributed to the model.

Why it matters

For teams evaluating image APIs, Craft is another production option alongside the models The Verge names from OpenAI, Microsoft, and Google. Its distinguishing claim is architectural: a planning stage that resolves spatial relationships, relative scale, visual hierarchy, and text information before pixels are generated, which Canva says produces legible text and precise object placement.

Canva says it used human-designed templates, posters, social posts, presentations, infographics, menus, and typography-rich layouts, rather than photographs, while also saying the model can create realistic scenes, digital art, and environmental imagery.

Output constraints are explicit: up to 2K resolution and more than 30 aspect ratios. TechCrunch reports images generated with Craft can be converted into layered designs using Canva Magic, and Canva says users can make precise edits without changing other parts of the image.

On integration, the model is reachable through the Canva API, and the company plans to let AI agents plug in through the open Model Context Protocol, with TechCrunch reporting planned support for image generation in third-party chatbots including ChatGPT and Gemini. Canva also says it is developing more efficient versions, Craft-fast and Craft-ultrafast, which TechCrunch describes as cost-efficient and names Craftfast and Craft-ultrafast.

TechCrunch notes Canva previously used a range of image models on its platform, and that the in-house model gives the company better quality and speed control in addition to cost effects.

What we don't know yet

Neither source gives latency figures, rate limits, or API terms for Craft, so throughput for production workloads is unknown. The Design Arena ranking is described as crowdsourced; how it maps to task-specific accuracy on legible text or object placement is not established, and Canva's precision claims are not independently tested in either source. Timing for Craft-fast and Craft-ultrafast is given only as "in the future" and "soon," and the two sources spell the fast variant differently. The MCP integration and third-party chatbot support are described as planned, with no stated availability date or supported client list. Licensing, commercial use rights, and how the template training data affects output ownership are not addressed. How Canva Magic's layer conversion behaves on Craft outputs in practice is also not detailed.

Sources: The Verge AI / TechCrunch AI

How this Deep Take is made: it is written with AI assistance from the linked sources, then checked against them sentence by sentence, and every fact links to where it came from.