Canva has built its own image generation model, which is a slightly awkward thing to announce when you have spent years stitching together other companies' models. The new one is called Canva Craft, it launched on Thursday, and its main selling point is that it thinks about a layout before it draws anything. Given how many AI images still come back with signage written in confident gibberish, that is a reasonable place to start.
What Canva Craft actually is
Canva Craft is the first image generation model from Canva's research team. It is aimed squarely at the established players, though the reports differ on exactly which ones (more on that below).
The output specs are straightforward. Images can reach up to 2K resolution, in roughly 30 aspect ratios meant to fit different platforms. Canva pitches it for the obvious design work, such as social posts and marketing content. It also says the model can produce realistic scenes, digital art and environmental imagery.
The two-stage trick
Most image models take a prompt and go straight to pixels. Craft inserts a planning step first.
- Stage one: the plan. The model reads the prompt and drafts a blueprint. That blueprint covers spatial relationships between elements, relative scale, visual hierarchy, and the information needed to generate text.
- Stage two: the image. The model then renders the picture from that plan.
Canva says the approach buys "greater precision, consistency and control." In practice, it claims accurate object details, legible text, and the ability to make precise edits without disturbing the rest of the image. Anyone who has asked an image model to fix one word on a poster and watched it redraw the entire poster will understand why that last claim matters.
Trained on designs, not photographs
The training data is the more distinctive choice. Rather than learning mainly from photographs, Craft was trained on what Canva describes as "millions of human-made templates" from its content library. That means posters, social posts, presentations, infographics, menus and typography-heavy layouts.
This is a sensible use of an asset Canva already owns. A model that has seen an enormous number of menus is probably better at menus than one that has seen mostly landscapes. It also fits Canva's broader product, because the output is meant to become an editable design, not a flat picture.
That connection runs through Canva Magic, the company's image-to-design transformer. Images made with Craft can be converted into layered designs with it. The layer-aware model behind this was released last year, and it is trained to understand designs and their layers. In April, Canva also added prompt-based editing, letting users create or alter a design by describing it, and brought layers to AI-generated images so one part can change while the rest stays put.
Researchers from Leonardo.ai, the media generation platform Canva acquired in 2024, contributed to the new model.
How it scores
Canva Craft currently sits fourth on Design Arena's leaderboard of AI image generation models. Design Arena is a crowdsourced benchmark for AI designs. The three models ahead of it are all OpenAI's: GPT-Image-2.5 Sunburst, GPT-Image-2.5 Flare and GPT-Image-2.
Fourth place on a crowdsourced board is respectable for a first attempt, and it is also a polite reminder that the model it is said to compete with is still ahead of it. The sources give no other benchmark figures, so that ranking is the only external yardstick available.
Where you can use it
Craft is live now for all Canva users. It is reached through the Dream Lab tab. One report describes the route as the three-dot option on the left-hand menu, then Dream Lab. It will also be offered through the Canva API.
Canva also plans to let outside AI tools use it through the open Model Context Protocol (MCP). One report says that this includes image generation inside third-party chatbots such as ChatGPT and Gemini. The other describes it more broadly as letting AI agents plug into Craft. Canva says the model ships with input and output moderation, privacy and safety controls, and reporting tools.
Two cheaper variants are in the works, with names given as Craftfast (or Craft-fast) and Craft-ultrafast. Canva says they will be more cost-efficient and will arrive soon, with no date attached.
The business angle, politely stated
Canva has been generating images with a mix of third-party models. Owning one gives it more say over quality and speed, and Canva named both as reasons. One report also reads the move as a likely way to save on API costs, which is an inference from the reporter, not a figure Canva published, though Canva did cite cost as a factor alongside quality and speed control.
There is a financial backdrop. Canva was reported to be on track for $4 billion in revenue last year. It then cut its growth forecast for 2026, and external analysts reportedly now put its valuation at $35 billion, down from $42 billion. Those valuation figures come from analysts, not from the company.
None of that proves the model was built to soothe a worried balance sheet. But a company that has just trimmed its forecast is not hurt by having a cheaper, in-house way to make the product it sells.
Where the reports disagree
The accounts of the launch differ in a few small ways.
| Detail | One account | The other account |
|---|---|---|
| Rival model makers | Google, OpenAI, ByteDance and Meta | OpenAI, Microsoft and Google |
| Aspect ratios | A possible 30 | More than 30 |
| Cheaper model name | Craftfast | Craft-fast |
The rival list is the largest gap, and neither account says where the list came from. The ratio count and the spelling look like loose reporting of the same announcement. I would treat "about 30" as the safe reading and wait for Canva's own documentation on the names.
What to make of it
The interesting bet here is not that Canva can out-draw OpenAI, since the leaderboard says it cannot yet. The bet is that a design platform can win by being good at the boring, commercially valuable things, like putting the right words in the right place on a menu.
If the plan-first approach works as described, the payoff is less time spent fixing spelling in a generated poster. If it does not, Canva still has a house model, two cheaper siblings on the way, and a fourth-place finish to improve on.