Canva has released Craft, its first in-house image generation model trained from scratch, and said all Canva accounts worldwide can use it in Dream Lab starting Oct. 22.
Developers can also access Craft through a public API, though Canva has not disclosed pricing or whether any free usage tier will be offered.
The company also said Craft will later be callable by AI assistants such as ChatGPT and Gemini through Canva’s MCP server, in addition to its website and API access points.
How Craft works
In its announcement, Canva said Craft splits image generation into two stages. The first is a planning model that reads the prompt and writes a composition plan covering which elements should appear, where they should be placed, their relative size, the visual hierarchy, and any text that should be included in the image.
The second is a drawing model that renders the image based on that plan. Canva said Craft supports output at up to 2K resolution, with more than 30 aspect ratios and custom sizing options.
Canva described the setup as similar to sketching first and coloring later. The company said the approach helps keep text clearer, improves object placement, and allows users to change one part of an image without affecting the rest.
Images generated by Craft can also be separated into editable layers through Magic Layers, removing the need to redraw them from scratch.
Language support and use cases
Canva said prompts can be written in any language, while text inside generated images supports 55 languages. As one example, the company said a request for a purple sky, long shadows, and a solitary tree would produce all three elements, with emphasis placed where the user intended.
The company said the model is suited for posters, social media posts, presentation visuals, portraits, landscapes, and illustrations.
Training data built around design templates
Canva said most image generation models are trained mainly on photos, while Craft was trained on millions of human-made designs from Canva’s template library. Those materials include posters, social media posts, presentations, infographics, restaurant menus, and layouts with heavy text.
According to Canva, the templates teach layout, hierarchy, and font pairing rather than what a finished output should look like.
The company said it then added supervised fine-tuning and reinforcement learning based on feedback from professional designers. In internal evaluations, designers and dozens of researchers compared more than 25,000 image pairs, and when automated scoring conflicted with designer judgment, the designer view took priority.
Why Canva built its own model
Cost is a central reason. Australian tech outlet Startup Daily, citing an August shareholder letter, reported that Canva CEO Melanie Perkins wrote that the company had become “overly reliant on frontier models,” referring to top-tier large models from providers such as OpenAI and Google, and that the average serving cost per AI task was too high.
Canva also cut its revenue growth forecast for the year from 30% to 20%. Revenue for the quarter ending in June was $921.9 million, up 25.2% year over year, below the target set earlier in the year.
The same report said Canva’s existing in-house models had reduced serving costs per AI task by 90%, with image costs 30 times lower than frontier models and video costs 17 times lower. Those figures refer to existing models such as Leonardo’s Lucid Origin, not Craft.
Craft’s own cost profile is still unclear
Canva said it has used model distillation to cut the number of compute steps needed for Craft image generation to a small fraction of the original level. It also built its own inference engine, which made the planning stage about an order of magnitude faster than the first internal version.
Even so, the company has not yet shown what Craft itself will cost at scale. Canva said that may become clearer after the launch of two lower-cost versions, Craft-fast and Craft-ultrafast, which it described as coming soon.
Scale makes small cost differences matter
Canva said it now serves more than 250 million users, generates more than 400 designs per second, and produces more than 1 billion designs each month. At that volume, even a difference of a few cents per image becomes significant.
Leonardo.ai, a Sydney-based generative AI startup acquired by Canva in 2024, contributed researchers to the Craft project. According to a Canva Research profile, the internal lab has more than 160 research scientists and engineers and sits within an AI division of about 400 people.
Valuation pressure is also in focus
Australian outlet SmartCompany reported that an external valuation commissioned by early Canva investors Blackbird and Airtree cut the company’s valuation by 17% to $34.9 billion, down from a 2025 peak of $42 billion.
Canva’s own annual employee share valuation also fell from $38.9 billion to about $31 billion.
Canva cites DesignArena ranking
Canva cited DesignArena rankings and said Craft debuted at No. 4 among all image generation models, which led the company to describe itself as the world’s second most advanced AI lab for image generation.
DesignArena is a design evaluation site that ranks models through crowd-sourced blind voting. In category results cited by Canva, Craft ranked No. 1 in portraits among 87 models and No. 1 in landscapes and nature among 85 models. It ranked No. 2 in abstract patterns and No. 4 in products.
The report also noted that these were results cited by Canva itself and reflected the model’s position shortly after entering the leaderboard, when the vote count was still limited and rankings could later change.
For now, Canva has not disclosed pricing, and neither the two cheaper versions nor MCP access has gone live. The currently available product is the base Craft model through Dream Lab and the public API.