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

What changed

Canva moved from relying on outside image models to shipping one built by its own research team. Craft splits the job in two: the first stage interprets a prompt and builds a plan covering spatial relationships, relative scale, visual hierarchy and text requirements; the second renders the image at up to 2K resolution across more than 30 aspect ratios. Canva says this produces legible text, accurate object detail and edits that leave the rest of an image untouched.

Training used designs rather than photographs, drawn from Canva's library of human-made templates including posters, social posts, presentations, infographics, menus and type-heavy layouts. TechCrunch reports researchers from Leonardo.ai, which Canva acquired in 2024, contributed to the model.

Who is affected

Every Canva user gains access immediately: the model sits in the Dream Lab tab, reached through the three-dot option on the left-hand menu. Canva also plans to expose Craft through its API and to third-party chatbots such as ChatGPT and Gemini via the Model Context Protocol, which would extend it to developers and agent builders.

In its announcement Canva cited input and output moderation, privacy and safety controls, and reporting tools. On Design Arena's crowdsourced leaderboard, Craft currently places fourth, behind three OpenAI GPT-Image models.

What is available

Craft is live now. Cheaper Craft-fast and Craft-ultrafast variants are described as future releases, not current products. TechCrunch notes Canva has previously drawn on multiple outside image models, and that in-house generation could reduce its API costs while giving it more control over quality and speed.