Canva Launches Canva Craft, Its First Image Generation Model

Canva has released Canva Craft, its first image generation model, and made it available to all users of its platform through the Dream Lab tab. TechCrunch reported the launch on October 22, 2026.

Canva's research team developed the model. Canva Craft generates images at up to 2K resolution and supports 30 aspect ratios, covering the formats Canva users typically move between, from social posts to presentation slides and print layouts.

According to TechCrunch, Canva Craft does not go straight from prompt to pixels. It first interprets the prompt and builds a plan covering spatial relationships, relative scale, visual hierarchy, and text generation. The image is then rendered against that plan. Text rendering has historically been one of the weaker areas for image models, so the explicit planning stage for typography and hierarchy is a deliberate design decision, not a cosmetic one.

Canva trained the model on human-designed templates from its own library instead of photographs. The training data therefore consists of layouts, type placement, and compositional choices made by professional designers, rather than the raw camera imagery that most general-purpose image models learn from. Researchers from Leonardo.ai also contributed to the model. Canva's 2024 newsroom post on Leonardo.ai described that company's models as able to generate hyper-realistic photos, character portraits, anime, and painterly styles.

Canva Craft is available to all users on the Canva platform through Dream Lab. Canva also says it will be available through the Canva API. In addition, Canva plans to support Canva Craft image generation inside third-party chatbots, including ChatGPT and Gemini, through Canva's MCP (Model Context Protocol, a standard that lets AI assistants connect to outside tools) server. That last item is a plan, not a shipped integration, and the timing has not been specified.

Images generated with Canva Craft can be converted into layered designs using Canva Magic, the feature set that connects generated imagery to Canva's editing environment. Canva's March 2026 announcement of Magic Layers stated that the feature is built on the Canva Design Model, which the company identifies as its foundation model. Canva also said Magic Layers is the first to generate designs that are editable from the start.

Canva is developing two cost-efficient variants, named Craftfast and Craft-ultrafast, for future release. Both names point to the usual trade-off in image generation between output quality and inference cost, the computing expense of producing each image. The existence of a faster tier suggests Canva expects high-volume use across its user base.

The launch sits inside a broader product push. In October 2025, Canva introduced its Creative Operating System and described the Canva Design Model as the world's first AI model trained to understand the full complexity of design. Bloomberg reported at the time that the foundational model enabled users to create and edit posters and short videos. In April 2026, Canva introduced Canva AI 2.0, an AI design platform that generates designs, documents, and websites and that the company describes as powered by that same foundation model. Bloomberg reported in April 2026 that Canva has 265 million monthly users.

The broader context here is the position of an image model inside a design tool rather than as a standalone generator. A standalone model competes on image quality and prompt adherence. A model that writes into a layered, editable document has to be judged on whether its output fits the rest of the workflow: brand kits, text that stays live and editable, and assets that can be reworked without regenerating from scratch. Canva's choices to train on template data, plan layout before rendering, and route output into layered files all point toward that second test.

In my view, the training decision deserves more attention than it usually gets in launch coverage. Templates encode layout conventions that professional designers have refined over decades of print and screen work. If the model has learned from those conventions, its default outputs should be easier to place into existing design systems. Whether that holds across the full range of prompts and styles will become clear only as users test the tool at scale.

The distribution plan is also worth watching. Reaching ChatGPT and Gemini users through Canva's MCP would make Canva's generation capabilities available inside assistants many people already use for drafting and research. Most Canva users are not professional designers, so that could bring design work into conversational interfaces. The integration is still a plan, and its details have not been published.

For developers, the most practical element is the API. Once Canva Craft is available through it, third-party products can build generation and layout features on the same model that powers Dream Lab. Teams already using Canva have a more immediate question: whether the 2K ceiling, the 30 aspect ratios, and the layered export path cover their production needs. They can now test that directly.