Canva Craft plans layouts before rendering images—and connects them to editable layers

Canva has introduced Craft, an image model that writes a composition plan before rendering the picture. The planning stage specifies elements, spatial relationships, scale, visual hierarchy and text—decisions that determine whether a poster or social graphic is usable as a design.[1]

Announced on 22 October, Craft is Canva Research’s first foundation image model trained from scratch. It supports output up to 2K resolution, more than 30 aspect ratios and text in 55 languages, according to the company. It is available through Canva’s Dream Lab and a public developer API.[1][3]

Why a picture might need a plan

Traditional text-to-image tools can impress with detail while struggling with the relationship between components. 'A red coffee cup to the left of a blue book, with the words MORNING NOTES at the top' is simultaneously an object-recognition challenge, a spatial-reasoning challenge and a typography challenge. A composition plan separates some of those decisions from the job of shading surfaces and rendering textures. The broad engineering idea is that an image's meaning includes its arrangement, not just its pixels.[1]

That is especially relevant to Canva's audience. A marketer may need a sale graphic with a discount amount in a specified corner; a teacher may want a classroom illustration with labels; a restaurant may need a promotional image with a readable dish name. In those situations, a photorealistic but incorrectly worded result is not a near-success. It may be unusable. A planning stage is a credible design choice because it gives the model a way to organise a request before dealing with the visual complexity of the finished asset.[1]

At launch, Canva reported Craft in fourth place among image-generation models on DesignArena, with Canva Research ranked second among the labs represented. These are two different rankings; the lab position was not a separate laboratory test of the model. The results are company-reported snapshots of the platform’s human-preference assessments.[1][3]

How the image becomes editable

Canva explicitly connects Craft to Magic Layers, which can transform generated images into editable design elements. The company says the model’s composition structure and JSON boundary boxing make that conversion more accurate and faster. The workflow therefore has two stages for a designer: generate the image, then use the layer-conversion capability when individual elements need editing.[1]

The developer announcement identifies the API model as craft_v1_2k. This offers a programmatic route to generated images outside the Canva editor. Teams building a publishing workflow should distinguish that image response from the separate editing experience available inside Canva.[2]

Where planning should make the difference

A useful comparison would focus on ordinary design errors: the wrong word, an object in the wrong place, or a layout that breaks when the aspect ratio changes. A multilingual poster adds another test—whether longer translated wording retains the intended visual hierarchy. These are proposed evaluation tasks, not tests performed for this article.[1]

The product brings model development, image generation and editing closer together within one design platform. That connection is commercially important: the time saved depends on how much of a first image a user can retain and correct, rather than how quickly it appears.[1][2]

Craft’s distinctive proposal is to settle the composition before rendering the pixels, then preserve enough structure to support editing. Exact wording, placement and successful conversion into layers will determine how useful that sequence is for everyday design.

Report an error or suggest a correction