Canva’s First Homegrown Image Model Debuts Fourth on DesignArena, Behind Only OpenAI

By Canva’s own count, OpenAI is now the only AI lab ahead of it in image generation. The claim rests on Canva Craft, the first image model built and trained from scratch by the company’s in-house lab, Canva Research, which went live for every Canva user on Thursday and debuted in fourth place on DesignArena, a crowdsourced leaderboard for AI design.

All three models above it belong to OpenAI: GPT-Image-2.5 Sunburst, GPT-Image-2.5 Flare and GPT-Image-2, The Verge reported. Canva said Craft ranks first out of 87 models in DesignArena’s People & Portraits category and first out of 85 in Landscapes & Nature. Adobe put those same OpenAI models inside Firefly on their launch day in September.

Why Canva wanted a model of its own

Canva’s AI features have run on a blend of its own models and partners’. Stef Corazza, Canva’s head of AI research, wrote in a post introducing the lab that the company needed “the flexibility to develop and run our own models alongside those from our partners.” He spelled out why the cost side counts at Canva’s size. People make more than one billion designs a month on the platform, and “even a variation of a few cents per generation makes an enormous difference.”

Corazza’s group has grown to more than 160 research scientists and engineers, part of a 400-person AI unit. It took in Leonardo.ai, which Canva bought in 2024, and TechCrunch reported that Leonardo’s researchers worked on Craft. Back at Canva Create in April, the company said its earlier Lucid Origin image model was 5x faster and 30x cheaper than comparable frontier alternatives, and that its first generation of models had taken more than two years to develop, against as little as a month for new ones. Two faster versions of the new model, Craft-fast and Craft-ultrafast, are in the works, and Canva said they are meant to deliver images “at a fraction of the cost of leading frontier models.” It gave no prices or dates for them.

It sketches the layout before it draws

Craft splits the job in two. A planning model reads the prompt and writes out the composition, down to where each element sits and what text has to appear, and a separate rendering model turns that plan into an image of up to 2K resolution in more than 30 aspect ratios. Most image models learn mainly from photographs, the company said. Canva trained Craft on millions of templates from its own library, including posters, presentations, menus and infographics, and its professional designers, working with dozens of researchers, compared more than 25,000 image pairs head to head against leading models against leading models. “When automated evaluations and human judgments differed, we sided with the designer,” the company wrote.

Canva is pitching its public API to product teams that build marketing platforms, e-commerce sites, ad builders and print services. Prompts can be written in any language, and Craft can render text in 55 of them. A link to AI assistants through the Canva MCP Server is listed as coming soon; once it arrives, the company said, an agent could make a launch graphic or a social post as one step in a bigger task.

Developers reach Craft by pinning the model value craft_v1_2k on the image generation endpoint of the Connect API, which Canva still labels a preview. Because Craft plans before it renders, a job takes roughly 50 to 60 seconds end to end, Canva said in its developer forum post, and each one spends the user’s AI credits. If Craft is unavailable, a pinned request fails with a model_not_available error, and dropping the pin lets Canva hand the job to one of its other models.