On October 22, 2026, Canva released its first in-house image generation model — Craft. The model was developed by the company's research team, Canva Research. According to TechCrunch, Craft competes directly with similar models from Google, OpenAI, ByteDance, and Meta. The official announcement was published on the company's newsroom page. According to Canva, switching to its own model helps save on API costs and gives better control over quality and speed.
How the model works
Craft uses a two-stage architecture that differs from traditional models: planning the composition is separated from rendering the image. In the first stage, a dedicated planning model interprets the user's prompt and writes a detailed plan of the composition. The plan captures which elements will appear, their spatial relationships, relative sizes, visual hierarchy, and information about text in the image. The planning model considers each element of the prompt individually: which objects will be in the scene, where they sit relative to one another, which are large and which are small, and where the visual emphasis lies — all of this is decided before generation. Text gets special attention: the content, placement, and size of the wording are an integral part of the plan. In other words, before it starts drawing, the model builds a complete map of the scene: where each object will stand, how its size compares to those around it, and how the text will look in each spot — all fixed in the plan in advance. In the second stage, a separate rendering model turns this plan into a high-quality image.
The model supports up to 30 different aspect ratios — from square social posts to wide banners — and generates images at up to 2K resolution. Planning the composition in advance improves accuracy in complex layouts: text comes out legible, objects are placed correctly, and visual relationships stay consistent. The architecture gives precise control over individual elements — part of the image can be edited selectively while the rest remains untouched.
What it was trained on
Instead of photographs, Craft was trained on templates created by human designers in the Canva library. The training set included professional posters, social media posts, presentation slides, infographics, menus, and typography-rich layouts. Each template is a finished decision by a designer: color harmony, white-space distribution, the balance of text and imagery. Rather than memorizing the appearance of finished images, such templates teach the model the principles behind effective design: how a layout is built, how hierarchy is established, and how typography blends with composition.
According to Canva, training on design templates delivers three practical results. First, fine details of objects come out accurately — the model learned the precise lines and shapes from the templates. Second, text in images is legible: one of the most important criteria for a design tool. Third, users can precisely edit an individual element without touching the rest of the image — for example, updating one piece of wording or moving one object without affecting the whole image. As a result, an image created by the model remains fully editable afterward.
Ecosystem integration
Craft is launching not as a standalone tool but as part of the Canva ecosystem. Images created with it can be turned into layered designs through Canva Magic — the company's image-to-design transformer. This means generation is not the final stop: the resulting image can be split into layers, with each element edited separately. Researchers from the Leonardo.ai team, which Canva acquired in 2024, also contributed to building the model.
The model is available to all users through the Dream Lab tab — from this tab in the interface, you can type a text prompt and create an image in Craft. For developers, Craft is also offered through the Canva API, which makes it possible to connect the model to external apps and services. The company plans to support image generation in third-party chatbots such as ChatGPT and Gemini via Canva MCP in the future — that integration is currently marked as "coming soon".
The company named two reasons for switching to its own model: saving on API costs and better control over quality and speed. Extending this logic, Canva announced two cheaper variants — Craftfast and Craft-ultrafast: they will be released soon and will allow generation at lower cost than the main model.
According to the company, the DesignArena ranking makes Canva Research the world's second most advanced AI lab in image generation with the launch of Craft:
According to the DesignArena ranking, Craft makes Canva Research the second most advanced AI lab in the world in image generation. — Canva newsroom, 10/22/2026
