Canva's first in-house image model, Craft, plans the layout before it draws and returns editable layers for design teams

The useful part of Canva's new image model is not the picture. It is that someone can still edit it.

What happened

Canva released Canva Craft, its first image generation model built in house by Canva Research. Researchers from Leonardo.ai, which Canva acquired in 2024, contributed.

Per TechCrunch and The Decoder:

• Craft works in two stages. It first plans spatial relationships, scale, visual hierarchy and text, then generates the image at up to 2K resolution in around 30 aspect ratios.
• It was trained on human-designed templates from Canva's library, such as posters, social posts, presentations, menus and typography-heavy layouts, rather than mainly on photos.
• Canva says it renders text in 55 languages and that results can be broken into editable layers.
• Canva says Craft placed fourth overall on the DesignArena ranking.
• It is available to all users in Dream Lab and, per The Decoder, through Canva's public API. Canva plans support in chatbots like ChatGPT and Gemini through its MCP.
• Two cheaper variants, Craft-fast and Craft-ultrafast, are coming.

My take

Most image steps I have built into content pipelines fail in the same place: the text. A post graphic with a misspelled offer or a mangled price gets rejected, and the whole image gets regenerated, again and again.

A model that plans the layout first and returns something a designer can open in layers changes the review step. The reviewer fixes one line instead of rerolling the image.

How I would test it before putting it in a client flow:

1. Check exactly what your client's Canva plan allows through the API before you design around it.
2. Run it against the client's real templates and brand text, including any non-English copy.
3. Keep a human approval step, and store which model made each asset. If those assets end up in ads, you may need that record for disclosure rules.
4. Benchmark the cheaper variants when they ship. For high volume social posts, speed and cost per image often matter more than top quality.

I would treat leaderboard rankings as a starting point and judge it on your own templates.