Canva launches Canva Craft, its first in-house AI image generation model, available to all users

Canva released Canva Craft, its first in-house image generation model, to all users worldwide through its Dream Lab tool. It works in two stages: it plans the design layout first, then generates the image. It outputs up to 2K resolution in more than 30 aspect ratios and was trained on millions of human-made templates from Canva's library. On Design Arena's crowdsourced leaderboard it ranks fourth, behind three OpenAI GPT-Image models. MCP access for AI agents and faster Craft-fast and Craft-ultrafast variants are promised but have no dates.

Canva now owns a model rather than depending only on outside providers. That could give it more control over cost, behaviour and design-specific features such as text and layout. Its large existing user base means many people will use it without choosing it. Fourth place on one leaderboard suggests it is competitive but not the best available, so the case for switching away from OpenAI models is weak on quality alone. Designers and marketing teams already in Canva have little new to decide, but rivals may feel pressure on bundled pricing.

Claims of greater precision, consistency and control, and of text legibility and object placement, come from Canva and are not independently shown. The Verge's framing of a direct fight with OpenAI, Microsoft and Google overstates things when Canva ranks behind three OpenAI models. A single crowdsourced leaderboard is a thin basis for judging quality, and the coverage gives no figures on usage limits, cost or training data licensing beyond the templates.

1. 01Whether Craft's Design Arena rank holds or rises as more votes come in, and how it compares with new OpenAI and Google image models over the next one to two months.
2. 02Whether Canva ships the promised MCP access and the Craft-fast and Craft-ultrafast models, and on what dates, pricing and usage limits.
3. 03Evidence of real adoption: Canva-reported Dream Lab usage numbers, any change in paid-plan limits, and independent tests of text rendering and layout accuracy against GPT-Image models.