Canva Craft: first image model plans before drawing
Canva's first in-house image generator drafts a layout plan first, then renders up to 2K across 30 aspect ratios, trained on templates.
In short
Canva Craft is the company's first in-house image generation model, and it runs in two stages: a planning pass fixes placement, scale and text before any pixels are rendered.
At a glance
- Two stages: Craft drafts placement, scale, visual hierarchy and text settings, then renders the picture.
- Output runs up to 2K resolution across 30 aspect ratios.
- Trained on human-made Canva templates: posters, social posts, presentations, infographics, menus.
- Live now for every Canva user in the Dream Lab tab and via the Canva API; MCP access for ChatGPT and Gemini is planned.
- Canva hit $4 billion in revenue in 2025; its valuation slid from $42 billion to $35 billion after a forecast cut.
Canva Craft is the company's first in-house image generation model, and it refuses to draw right away: a planning stage settles placement, relative scale, visual hierarchy and text before the renderer touches a pixel.
A plan before the pixels
Most generators map a prompt straight onto an image. Per TechCrunch, Craft inserts a deliberation step that works out what sits where, how big each element is relative to the rest, and how any lettering should be laid in. The second stage then produces output at up to 2K resolution in 30 aspect ratios.
The payoff Canva is aiming for is the part diffusion models usually fumble: object detail that holds up and type you can actually read. Edits are meant to stay local too, so changing one region does not reroll the whole frame.
Trained on layouts, not landscapes
Instead of leaning mainly on photographs, the training set draws on human-designed templates from Canva's own library: posters, social posts, presentations, infographics, menus and type-heavy layouts. That choice points the model at a different target. It is optimized for a usable design, not a convincing photo.
Some of the research muscle arrived with the Leonardo.ai acquisition in July 2024. Canva Magic, shipped in October 2025, can also break a generated image back into editable layers.
Where you can run it today
Craft is available immediately to all Canva users through the Dream Lab tab, and it can be called through the Canva API. Canva's MCP connector is meant to open the model to outside chatbots including ChatGPT and Gemini. Two cheaper variants, Craftfast and Craft-ultrafast, are on the roadmap with no date attached.
Into a crowded field
TechCrunch lists Google, OpenAI, ByteDance and Meta as the direct competition. The same report sets out the business backdrop: Canva reached $4 billion in revenue in 2025, then trimmed its 2026 growth forecast, and its valuation fell from $42 billion to $35 billion.
What we could not check
A second write-up of the launch at The Verge could not be retrieved during reporting. Its claim that Craft also squares off against Microsoft therefore stays unconfirmed here. We likewise have no independently verified benchmark scores, no pricing for the planned budget variants, and no detail on the rights status of the templates used in training.
FAQ
What is Canva Craft?
Canva Craft is Canva's first in-house image generation model. It plans a composition first, then renders images at up to 2K resolution in 30 aspect ratios.
How is Craft different from other AI image generators?
It splits the job in two. Stage one sets placement, relative scale, visual hierarchy and text parameters; stage two renders from that plan, which is meant to improve object detail and legible lettering.
Can I use Canva Craft from ChatGPT or Gemini?
Canva says third-party chatbot access, including ChatGPT and Gemini, is planned through its MCP connector. Today Craft is reachable in the Dream Lab tab and through the Canva API.
