Canva ships its first in-house image model to take on Google and OpenAI

Design platform Canva released Canva Craft on October 22, its first image generation model built by its own research team and aimed squarely at comparable models from Google and OpenAI. TechCrunch and the Australian Financial Review both framed the launch as Canva racing to keep pace in generative imagery.

The model works in two stages. It first reads the user's prompt and produces a plan covering the spatial relationships of elements, relative scale, visual hierarchy and the information needed for text; it then renders the image from that plan, with several aspect ratios to fit different platforms. Canva says the split lets the model reproduce object detail more accurately, render legible text, and allow precise edits to one part of an image without disturbing the rest.

The training data is the most distinctive choice. Rather than leaning on photographs, Canva trained the model on design — human-made templates from its library, including posters, social posts, presentations, infographics, menus and typography-heavy layouts. That closes a loop across the product: an output can be turned back into an editable, layered design through Canva Magic, the company's image-to-design transformer.

Canva is making the model available to every user through the Dream Lab tab inside its platform, and to developers through the Canva API. It also plans to support image generation inside third-party chatbots such as ChatGPT and Gemini through Canva's MCP.

The commercial logic is straightforward. Canva has long mixed several outside image models inside its product; owning one turns image generation from a recurring API bill into an internal capability it can tune. Owning the model matters less for bragging rights than for unit economics: it lets Canva set the cost curve of generation itself, with finer control over quality and speed. Researchers from Leonardo.ai, which Canva acquired earlier, contributed to the new model.

Canva says it is also developing two cheaper variants, Craftfast and Craft-ultrafast, to be released soon. That fits its user base: with a large population of free and light users, a model that stays open to everyone only makes financial sense once inference costs come down.

Seen competitively, the shift is clear. The moat for design tools used to be templates, collaboration and brand assets; it is moving toward generation quality and cost per image, while foundation model companies push generation straight into chat boxes and bypass the design layer altogether. The AFR headline summed it up as Canva trying to outrun Google and OpenAI, which makes this as much a distribution fight as a product one.

What to watch: whether a design-native model proves steadier than general models on real commercial assets, whether the chatbot integrations feel natural to use, and whether the low-cost versions can carry free-tier volume. The answers will decide if the in-house model is a defensive patch or the next growth engine.