Design platform Canva has released its first image generation model, a system called Canva Craft built by the company's in-house research team. The launch, reported by TechCrunch on Thursday, puts the Australian design giant in direct competition with image models from Google, OpenAI, ByteDance and Meta — and marks a shift from Canva's long-standing strategy of renting image generation from other providers.
A design-first image model For more context on this story, see our ongoing AI news.
Most consumer image generators were built to produce striking one-off images. Canva Craft was built for something narrower and, for Canva's business, arguably more valuable: producing usable design assets. The company says the model was trained not on photographs but on human-designed templates drawn from Canva's own library — posters, social posts, presentations, infographics, menus and typography-rich layouts.
That training choice shows up in what the model emphasizes. Canva says Craft generates images with accurate object details and legible text, two of the most common failure points for image models used in marketing and communications work, where a garbled headline ruins an otherwise good graphic.
How Canva Craft works
The model operates in two stages, according to Canva. In the first stage, it interprets the user's prompt and builds a plan: the spatial relationships between elements, their relative scale, the visual hierarchy, and the information needed for any text that will appear in the image. Only in the second stage does it generate the image itself, at up to 2K resolution and in as many as 30 different aspect ratios for different platforms.
Precision editing without collateral damage
The planning stage also enables what Canva describes as precise edits: users can change one part of a generated image without the model regenerating — and silently altering — everything else. For anyone who has watched a carefully art-directed graphic mutate after a single revision request, that distinction matters.
Because the model was trained on layered design templates rather than flat photos, images generated with Craft can be converted into fully layered designs using Canva Magic, the company's image-to-design transformer released last year. That pipeline — prompt to image to editable, layered design — is the clearest expression of what Canva thinks an image model should do.
Available to all users, with API access next
Canva Craft is rolling out to all Canva users through the Dream Lab tab, and the company says it will also be exposed through the Canva API. Plans are in place to support image generation inside third-party chatbots such as ChatGPT and Gemini via Canva's MCP server, which would let users generate on-brand designs without opening Canva at all.
The company is also developing two cost-optimized variants, Craftfast and Craft-ultrafast, which it says will be released soon.
Why Canva built its own model
Until now, Canva has relied on a rotating cast of third-party image models inside its platform. Running its own model gives the company three things it did not have before: lower API costs, control over quality, and control over speed. At Canva's scale, the economics alone could justify the research investment — every generated graphic that moves from a paid external API to in-house infrastructure drops straight to the bottom line.
The launch also leans on acquired talent. Canva bought media generation platform Leonardo.ai in 2024, and the company says researchers from Leonardo contributed to the new model. Craft is, in effect, the first major consumer-facing product of that acquisition.
The competitive picture
Canva is entering a crowded field. Google, OpenAI, ByteDance and Meta all operate frontier-scale image generators, and each has folded image generation into chat products with enormous distribution. Canva's bet is that design-native output — editable layers, correct text, platform-ready aspect ratios — matters more to its users than raw model generality.
There is also a defensive logic. If image generation becomes a commodity, whoever owns the workflow — the templates, the brand kits, the export pipeline — captures the value. Craft is Canva's attempt to make sure the model layer serves the workflow layer, rather than the other way around.
A high-stakes moment for Canva
The launch lands at a delicate time for the company. Canva was on track for $4 billion in revenue last year, according to TechCrunch, but the company cut its growth forecast for 2026, and external analysts have reportedly marked its valuation down from $42 billion to $35 billion.
Owning the generation layer does not solve a growth problem by itself. But it tightens the feedback loop between what Canva's hundreds of millions of design sessions show about user intent and what its model learns to produce — a loop that competitors renting models from OpenAI or Google cannot easily replicate.
Whether Canva Craft can stand up to the marketing muscle of frontier labs is an open question. What is not in question is the strategy: Canva has decided that in the next phase of generative design, the companies that win will be the ones that own their models.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
