Canva Craft is Canva's own image model, and the API calls it craft_v1_2k behind a preview warning

Canva has launched Craft, its first image model, trained from scratch by Canva Research. It is available to all Canva users and through Canva’s public API.

How it works

Canva describes two stages. A planning model first writes a plan for the composition, then a rendering model draws it at up to 2K resolution in more than 30 aspect ratios. Prompts can be in any language, and Canva says Craft renders text in 55 languages, which is the hard part for most image models. It was trained with millions of Canva templates, so its strength should be designed layouts such as posters, slides and infographics rather than photographs.

Canva says Craft debuted fourth on the DesignArena leaderboard and calls itself the second most advanced image-generation lab on those rankings. Those are Canva’s claims; we have not tested output quality.

The API

In the Connect API spec, Craft appears as craft_v1_2k (“CRAFT-1 2K”) on POST /v1/image-generations, next to lucid_origin and z_image_turbo. Canva recommends leaving the model unset so it can pick one and fall back if needed. Three details matter before you build:

1. Preview. The endpoint is marked preview: breaking changes can arrive without notice, and public apps that use it will not pass Canva’s review.
2. Credits. Each job spends the user’s AI credit allowance, and an empty allowance returns 429 credit_quota_exceeded.
3. Fewer shapes. The API lists six aspect presets, from square to 4:5, not the 30 plus the app offers.

Canva says MCP support is coming soon, along with faster Craft variants.

Why it matters

Canva has relied on partner models until now. Owning one tuned on its own templates gives it control over cost and lettering, which is what most Canva users actually need from image generation.