On Oct 22 2026 Canva released Craft, its first image model. Canva says it was built and trained from scratch by Canva Research, its in-house AI lab (Canva Newsroom, also covered by TechCrunch and The Verge).
It is live now for all Canva users globally, inside Dream Lab. If you build tools, it is also in Canva’s public API: the Canva Developers post lists the model as craft_v1_2k, marks it as a preview API, and says an image takes about 50 to 60 seconds and uses AI credits.
For ad work, Dream Lab is where I would start.
Craft works in two stages: it plans the composition first, then renders. In the plan it works out the elements, the layout and how they sit next to each other, the scale, the hierarchy and the text (Canva, TechCrunch, The Verge).
Canva says that gives you legible text and precise placement. That is exactly where ad statics usually break: a headline you can’t read, a product in the wrong spot, or three things all shouting at the same size.
When the model plans hierarchy before it draws, a clear brief from you actually lands.
I don’t prompt it. I brief it like a director. Every brief I write has six parts: the product (what it is and what it looks like), the audience (who stops scrolling for this), the headline text (the one line the ad needs, word for word), the hierarchy (what is biggest, what is secondary, where the headline sits), the mood (the feeling, the light, the colors) and the aspect ratio (where it will run first).
Because Craft plans the layout before it renders, the hierarchy part is the one that pays off most. The ready-made version of this brief is in the members pack at the bottom.
Keep it short and quote it exactly. Put the headline in quotes in your brief, exactly as it should read, and keep it to one short line. One clear line beats a paragraph on an ad static every time.
Canva says you can prompt in any language and get in-image text in 55 languages (Canva Newsroom, Canva Developers), which matters if you make ads for brands selling in more than one market.
Still read every word before it goes anywhere (more on that in step 07).
Fix one thing at a time. Canva says Craft can make targeted changes without altering everything else in the image (Canva Newsroom; TechCrunch describes it as precise edits without changing other parts).
So when a draft is 90% there, don’t re-roll the whole thing. Ask for one change: move the product left, make the headline bigger, swap the background color.
Keep the rest. That is how I would work with a designer too, and it keeps a good composition from getting lost.
Pick your placements, then resize. Craft supports more than 30 aspect ratios and up to 2K resolution (Canva Newsroom, The Verge), so the same layout can go to feed, story and banner.
Canva also says Craft outputs are particularly well suited to Magic Layers, which can turn a generated image into fully editable design elements (Canva Newsroom).
For client work that is the big one: the winning static becomes layers you can still move, retype and adjust, instead of a flat picture.
The tool got smarter. Your direction is still the job. Before anything goes to a client I check three things.
Never let a model generate a brand’s logo. If any of the three is off, fix it with a targeted edit or in the editable layers before you send.
Two things Canva has announced as coming soon. The Canva MCP, so AI agents like ChatGPT and Gemini can work with Canva (Canva Newsroom, TechCrunch).
And Craft-fast and Craft-ultrafast, cheaper versions of the model (Canva says “at a fraction of the cost”; The Verge).
Neither is out yet, so I would not build a client workflow around them today. For now: Dream Lab, a director brief, targeted edits, Magic Layers, and a human quality check.
This guide is the full process. The ready-made version lives inside AI Creative OS on Skool, next to every other prompt pack I post.
