Canva launches Craft, its first in-house AI image generation model

Canva's first in-house model goes live

Canva has released its first internally developed image generation model, called Canva Craft. According to TechCrunch, the model was built by the design platform's own research team with the explicit goal of competing with image generators from Google, OpenAI, ByteDance and Meta.

A two-stage approach to generation

Craft works in two stages rather than producing an image in a single pass. TechCrunch reports that the first stage interprets the user's prompt and builds a plan covering the spatial relationships between elements, their relative scale, the visual hierarchy of the composition and the information needed for text generation. The second stage turns that plan into an image of up to 2K resolution, in any of 30 aspect ratios suited to different platforms.

Canva claims this planning step yields images with more accurate object details and legible text, a longstanding weak spot for image generators, and allows users to make precise edits without altering other parts of the picture.

Trained on designs, not photographs

The training data is what most clearly separates Craft from the rest of the field. Instead of photographs, Canva trained the model on human-designed templates from its own library, including posters, social posts, presentations, infographics, menus and typography-heavy layouts.

That choice has a practical consequence: images generated with Craft can be converted into layered, editable designs using Canva Magic, the image-to-design transformer the company released last year. Craft is therefore less a general-purpose art generator and more of a front door into Canva's core editing product.

Where it will be available

Craft is already available to all Canva users through the Dream Lab tab on the platform. Canva also plans to offer the model through its API, and to support image generation inside third-party chatbots such as ChatGPT and Gemini via the company's MCP integration.

Cost is part of the calculation

Until now Canva has relied on a range of external image models to power generation features on its platform, incurring API costs along the way. TechCrunch reports that bringing generation in-house is likely to reduce those costs, and Canva says ownership also gives it better control over quality and speed. Two cheaper variants, Craftfast and Craft-ultrafast, are in development and expected to arrive soon.

The project also draws on Canva's 2024 acquisition of the media generation platform Leonardo.ai. The company says researchers from that team contributed to the new model.

Why it matters

The release lands at an awkward moment for Canva's finances. According to TechCrunch, the company was on track for $4 billion in revenue last year but has since cut its growth forecast for 2026, and external analysts reportedly estimate that its valuation has fallen from $42 billion to $35 billion. Owning the generation stack, rather than renting it from OpenAI or Google, is one of the few levers a platform of Canva's scale can pull to protect margins.

The training strategy is the more interesting bet. Legible text and editable layers matter far more to Canva's audience of marketers, students and small businesses than photorealism does, and generalist models from the big labs handle both poorly. By training on templates and routing the output straight into a layered design editor, Canva is competing on workflow integration rather than raw image quality, a fight where a design platform arguably starts with home advantage.