Canva announces the launch of Canva Craft, an image generation model developed by Canva Research, capable of producing visuals “in 2K resolution and in any format”. Unveiled on Thursday 22 October 2026, it is already available in the Dream Lab tab of the platform.
A model that plans before drawing
Craft is presented as the first foundation model, i.e., a generalist model serving as a base for various applications, designed and trained entirely by the company’s internal AI lab. Researchers from Leonardo.ai, the image‑generation platform acquired by Canva in 2024, contributed to its development.
Its particularity, according to the company, lies in its working method. Rather than producing a result immediately after receiving an instruction, called a prompt, the model separates “the planning of the composition and the rendering of the image”. A first module interprets the request and drafts a detailed plan that lists “the elements that appear, their spatial relationships, visual hierarchy and the texts that should appear in the image”. A second module then transforms this draft into an image.
Canva states that this approach enables the generation of complex layouts “with greater precision: readable text, exact placement of objects, coherent visual relationships”. It would also allow retouching a specific area without altering the rest of the visual. Moreover, the produced images adapt better to Magic Layers, a feature rolled out in March that allows reworking a fixed image layer by layer.
On the DesignArena benchmark, a comparative model test, Canva Craft is outperformed only by OpenAI models, the company says.
Training based on templates
Craft was trained on millions of templates, these layout models created by users and sourced from Canva’s library. This source, the press release says, distinguishes it from competing models “mostly trained on photographs”.
After this pre‑training phase, the team refined the model through supervised fine‑tuning, i.e., adjustment based on annotated examples, followed by reinforcement learning. These steps incorporated “feedback from Canva’s professional designers” to improve “instruction compliance, typographic accuracy and overall visual quality”.
Craft supports more than 30 image formats and understands prompts written in over 55 languages, including French. Canva presents it as a versatile tool, suitable for social‑media content, marketing assets such as posters or flyers, photorealistic images of people or places, as well as illustration, in flat, vector, pictorial or hand‑drawn styles.
Availability and upcoming versions
Canva Craft is available worldwide. It is integrated into Dream Lab, Canva’s AI‑powered creation space, accessible in the lower‑left corner of the interface. Developers can also integrate it into their own tools via the API, a programming interface that allows software to communicate with each other. The model will soon be offered on Canva’s MCP server, an access point that allows artificial‑intelligence agents to use its tools, “so that agents can create images ready to be embedded in a design”, the press release details.
The platform warns that “this first version of Craft is just the beginning”. Two faster variants, dubbed Craft‑fast and Craft‑ultrafast, are in preparation. According to Canva, they should “provide high‑quality image generation for a fraction of the cost of leading flagship models”. Their release date has not been disclosed.
