The Pulse
Canva Releases Craft, Its First In-House Image Model
Canva launched Craft, its first image-generation model built and trained by its in-house research lab, on October 22. The model plans image composition before rendering and is available globally in Canva and through its public API.
AI.info Team · Written with AI · fact-checked ·
Canva announced Canva Craft on October 22 as its first image-generation model developed and trained by its in-house AI lab. The company says the model is built for design work: it plans an image’s layout before rendering it, aiming to improve the placement of objects and text. Canva makes Craft available to users worldwide and to developers through its public API.
“Today, we’re unveiling Canva Craft, our first state-of-the-art image generation model, developed and trained by Canva Research, our in-house AI lab,” the Canva Team writes in the announcement. The launch puts a company-built model inside the same design platform where users can refine and reuse their creations.
Craft plans a composition before drawing it
Craft separates image generation into two stages. First, a planning model interprets a prompt and outlines the elements to include, their positions and relative sizes, the visual hierarchy, and any text. A separate rendering model then turns that plan into an image at up to 2K resolution.
Canva says the architecture supports more than 30 aspect ratios and lets users make targeted changes without unnecessarily altering other parts of an image. The company also says Craft can generate text in 55 languages. Those capabilities address a familiar design task: getting a generated image to fit a specific format while keeping its layout and wording usable.
The model’s outputs are designed to work with Magic Layers, Canva’s tool for turning generated images into editable design elements. That connection gives users a route from a generated image to components they can adjust inside a design, rather than treating the image as a finished, uneditable asset.
Canva trained Craft on its design library
Rather than centering its training on photographs, Canva says Craft’s training includes millions of human-made templates from its content library. The examples span posters, social posts, presentations, infographics, menus and layouts rich in typography. Canva says the templates teach the model about layout, hierarchy and how type fits into a design.
After pre-training, the company used supervised fine-tuning and reinforcement learning, with feedback from Canva’s professional designers. For internal comparisons, those designers and dozens of researchers assessed more than 25,000 pairs of images against leading models, according to Canva. The company says it favored designers’ judgments when they differed from automated evaluations.
Canva also touts Craft’s early DesignArena results, saying it debuted fourth overall and ranked first in the People & Portraits and Landscapes & Nature categories. Those standings are claims in Canva’s announcement, not a substitute for independent comparisons across real design tasks.
Global access starts; faster versions are planned
Canva says Craft is available now to all users globally through Dream Lab, as well as through its public API for developers and product teams. The company describes the API as a way to incorporate the model into tools such as marketing platforms, e-commerce sites, advertising builders and print services. Canva MCP access for AI assistants is marked as coming soon, rather than part of the current release.
The company is also developing Craft-fast and Craft-ultrafast, which it says will target lower generation costs. Canva describes its work on model distillation and a custom inference engine as ways to reduce the time and computation involved in generation, but it has not provided release dates or price figures for the variants.
For now, the concrete change is that Canva’s first model built from scratch by Canva Research is live in its design product and public API. Its stated focus is not just producing an image, but planning the composition and text that users may want to edit afterward.