Canva Craft plans the layout before rendering an image

Canva Craft begins an image with a plan: which elements belong in it, where they sit, how large they are and what text they contain. A second model renders that composition. Launched on 22 October in Dream Lab and the developer API, Canva’s first in-house image model targets a familiar design frustration—an attractive generated image that puts the important elements in the wrong places.[1]

The system supports images up to 2K resolution in more than thirty aspect ratios. The planning stage interprets visual hierarchy and required lettering as well as the relative sizes and positions of objects, giving the rendering model a more structured brief than an undifferentiated text prompt.[2][1]

Why planning first changes the problem

Traditional text-to-image experiences often tempt a user to submit a long paragraph, inspect a plausible image, notice two errors, and generate an entirely new version that fixes one problem while changing five things that were already right. By making composition an explicit intermediate representation, Canva is attempting to give image generation a structure closer to design work: objects have intended places, text has a role, and relationships should persist through revisions.[1]

That matters for a newspaper graphic, a menu, a promotional poster or a product card. The image's credibility may depend on accurate words and stable placement as much as artistic beauty. Canva says the structure also helps its Magic Layers technology separate generated material into editable elements. Such editability should be tested on realistic assets. A flattened AI picture that looks editable in a demo is not the same as a professionally usable layered design where every headline, illustration and background can be manipulated independently.[1]

Canva says it trained Craft using millions of human-created templates from its design library, including presentations, menus, posters and infographics. Those materials emphasise typography and layout as well as photographic appearance. The resulting model is intended to learn how visual elements work together in an artefact that has a purpose, such as making a heading legible or keeping a product prominent.[1]

The leaderboard needs a legend

Canva reports strong positions in DesignArena’s crowdsourced rankings, including first place in the portrait and landscape categories at launch. Those rankings measure preferences over the outputs presented to evaluators. A production design has additional requirements, including exact wording, accurate brand details and changes that preserve the parts already approved. Those are useful dimensions for a creator to test in their own workflow.[1]

For developers, a documented API is more concrete than a marketing rank. Canva's October announcement introduced the model identifier craft_v1_2k for its image-generation endpoint. The developer community notes that jobs are asynchronous, rely on user credits and may take around 50–60 seconds to complete, so applications should poll sensibly rather than assume near-instant results. Those timing details are operational observations for the current release, not guarantees of future performance.[1]

The revisions after the first image

A practical comparison would keep the same brief through several changes: move the heading, alter the product position, correct a line of text and export the result. That would reveal whether the planned composition helps preserve approved elements while making targeted edits. It is a proposed evaluation, not a test performed for this article.[2][1]

Craft’s promise is strongest in work where the first image is the beginning of an edit. A poster, product card or menu needs stable relationships between text and objects, then a way to refine individual choices. Planning composition before rendering gives Canva a route towards that workflow; the practical measure is how much revision the creator still has to do.[1]

Report an error or suggest a correction