Canva Launches Craft, Its First In-House Image Generator

Craft, the design platform's first self-built image-generation model, is now live for every user through the Dream Lab tab, the company said in a release reported by TechCrunch on Oct. 22. The tool is positioned to compete with image generators from Google, OpenAI, ByteDance and Meta, according to the report.

How does the two-stage generation process work?

TechCrunch reported the model interprets a prompt in two steps. First, it drafts a plan covering spatial relationships between elements, relative scale, visual hierarchy and any text that needs to appear. Second, it renders the image at up to 2K resolution across 30 possible aspect ratios sized for different platforms, per the report. That structure, the company told TechCrunch, is meant to keep object details accurate and text legible, and to let users edit one part of an image without disturbing the rest.

Why train on templates instead of photographs?

Rather than training on photographs, the platform said it built the model on human-designed templates from its own library — posters, social posts, presentations, infographics, menus and typography-heavy layouts — according to TechCrunch. That choice means the model's output can be converted into layered, editable designs using the platform's existing image-to-design transformer, released last year, the report said. Researchers from Leonardo.ai, the media-generation company acquired in 2024, contributed to the new model, per TechCrunch.

Who can access it, and through what channels?

Beyond the Dream Lab tab, the model will be available through the company's API and, soon, inside third-party chatbots such as ChatGPT and Gemini via an MCP integration, TechCrunch reported. Two lighter versions, called Craftfast and Craft-ultrafast, are in development for later release, according to the report. The push to embed generation directly into chat interfaces mirrors a broader industry trend, also visible in Meta's new AI-focused subscription plans.

What does this mean for the company's finances?

• $4 billion: approximate revenue last year, per TechCrunch.
• $42B to $35B: estimated change in the company's valuation after it cut its 2026 growth forecast, according to external analysts cited by TechCrunch.

Building the model in-house, rather than paying outside providers for every generation request, is likely to cut API costs tied to running multiple third-party models, TechCrunch reported, while giving the company more control over quality and speed.

What to watch

• Whether Craftfast and Craft-ultrafast ship, and at what price.
• How the MCP rollout into ChatGPT and Gemini performs once live.
• Whether in-house generation measurably lowers third-party API spend.
• How Google, OpenAI, ByteDance and Meta respond with competing releases.

Glossary

• Dream Lab: the in-app tab where users reach the platform's generative AI tools.
• MCP: Model Context Protocol, a standard letting outside apps like ChatGPT call a platform's tools directly.
• Image-to-design transformer: the existing tool that converts a flat image into a layered, editable design file.

For another example of creative tools built on unconventional training approaches, see HTT News' report on a San Francisco project using fly neural data to power creative AI.

For a Bay Area-baked gift, Stirred, Not Shaken ships in the U.S.