Canva Craft: the image is now made inside the design tool
Canva just dropped its own image model, trained on templates instead of photos, with results that can be split into layers. What that means for marketing teams, and what stays with a studio.
Miraki MindAI-assisted editorial · 23 October 2026 · 6 min read
On 22 October 2026, Canva introduced Canva Craft, the company's first in-house image model. It is built for design, not for photos: posters, social posts, presentations, menus. Images come out at up to 2K, can then be split into editable layers, and developers can reach it through Canva's API.
Our work starts earlier. We build the world first, then we shoot in it: characters, places and light are made as sheets and approved before a single shot is made. On Rapido, our spec ad for Electrolux, the vacuum was locked once as a three-view sheet and given to every product clip as the first reference. We built the flat first, so every cut opens onto the same flat.
Craft solves a different task: a strong single image in a layout, fast, in the tool a team already has open. That's not a small thing. It just moves the line between what a team does itself and what it brings a studio in for.
What is actually new
According to Canva, as reported by TechCrunch, Craft works in two steps. First, one model plans the composition: where each element sits, how big it is, what catches the eye first and where the text goes. Then a second model renders the image, at up to 2K and in many aspect ratios. Canva says Craft was trained on human-designed templates from its own library, not on photos. Researchers from Leonardo.ai, which Canva acquired in 2024, also worked on it. Inside Canva, Craft lives in Dream Lab, for all users.
The most interesting part is taking it apart. With Magic Layers, Canva's tool that turns an image into a design, the result becomes layers you can change one by one. So not a flat bitmap, but a layout you can keep working on. Canva says this allows targeted edits without touching the rest of the image.
The API is there too. In Canva's developer docs the model is called craft_v1_2k ("CRAFT-1 2K"), with two other models to choose from. The docs list the API as a preview: changes can land without a new version, and public apps that use it won't pass Canva's review. Every generation draws on the AI allowance of the person who triggers it, and the download link is valid for 24 hours. Announced, but without a date: Craft in ChatGPT and Gemini through Canva's MCP, and two cheaper variants, Craftfast and Craft-ultrafast.
What it means for a production
For key visuals, social formats and variants in many sizes, a fair amount is likely to move into the design tool. If you already lay things out in Canva, you make the image in the same place and adjust it there, without the export-import detour. For many teams that's simply practical.
A film is a different thing. On Pocobello, the character Oswald took seven attempts before his hair, glasses and posture were right. Every character was locked as a three-view sheet before any shot was made, and only then was the video model allowed to animate what happens in the frame. Face, costume and colour stay under human control across the whole film, not the model's. That work doesn't disappear just because a single image gets made faster. A strong image is one task. Keeping a character, a product and a flat the same across 45 seconds or four minutes is another.
Craft makes images, not motion. For us it might play a role early in the process, for moodboards or layout ideas for a brief. We haven't tested Craft yet.
What it means for a brand
Three things are worth a close look.
Rights: that Craft is trained on templates instead of photos is a clear statement about where it comes from. But the sources we could read say nothing about who owns the results or how they may be used commercially; the API docs are silent on it. Before a Craft image goes into a paid campaign, read the terms of your Canva plan.
Brand consistency: none of the sources we could read say whether Craft takes stored brand colours, fonts and logos into account when generating. A model that learned from other people's templates doesn't know your brand by itself. For us, consistency isn't a prompt but an approved sheet: the product, the character, the room, signed off once and then part of every shot as a reference.
Labelling: an AI image stays an AI image, even after it has been split into layers and retouched by hand. For us, labelling under the EU AI Act and cleared usage rights are part of the process. Whether a film needs a label, the AI Act Navigator in our Lab gives you a first assessment.
Our take and the limits
Canva Craft is a serious step: an image model built for layouts that delivers results you can keep working on. For teams already working in Canva, a lot that used to go outside gets faster. That's fine. What stays with a studio is what goes beyond a single image: characters and worlds approved before the first shot, a look that holds across a whole film, and rights, consent and labelling cleared before delivery.
A few things are still open. How well does Craft hold a brand across a whole campaign? What applies to rights and usage? What does an image cost through the API, and when does the preview become a stable API? That's what we want to know before we use Craft on a project.