Canva launches Craft AI model trained on templates to output editable layers
Canva built an AI image model trained on human designs instead of photos, turning prompts into editable layered templates.
Models · Source: TechCrunch
What happened
Canva just released Craft. It is the company's first in-house AI image generation model. The model works in two distinct steps to ensure accuracy. First, it acts as a planner to map out the layout. It determines spatial relationships, relative scale, visual hierarchy, and text placement. Then, a second model draws the final picture. It generates images in up to 2K resolution across 30 different aspect ratios. This split approach keeps text readable and puts objects exactly where they belong.
The training data sets Craft apart from competitors like OpenAI and Google. Instead of scraping internet photographs, Canva trained Craft on millions of human-designed templates from its own library. This massive dataset includes posters, menus, infographics, social posts, and typography-rich layouts. Because it understands design structure, the output pairs with Canva Magic to convert generated images into editable layers. Users can tweak text or move specific objects without altering the rest of the image.
Canva Research built the model using talent from its 2024 acquisition of Leonardo.ai and Kaleido. The research lab now employs over 160 AI engineers. Craft is available now to all users in the Dream Lab tab. Developers can also access it through the Canva Connect API. To cut down on massive generation costs, the company plans to release cheaper, distilled versions called Craft-fast and Craft-ultrafast soon. Canva will also support image generation in third-party chatbots like ChatGPT and Gemini through its MCP server.
Key facts
- 250 million — Canva users generating over 400 designs every second
- 1330 — Craft's Elo rating on DesignArena, sitting behind three OpenAI models
- 55 — Number of languages Craft can generate as legible on-image text
- $35 billion — Canva's reported valuation, down from $42 billion after cutting growth forecasts
Why it matters
Most AI image generators spit out flat pixels. If you want to change a typo or move a logo, you have to reroll the entire prompt and hope for the best. Craft changes this dynamic entirely by treating image generation as a structured design task. By outputting layouts that split into editable layers, builders can integrate AI directly into actual production workflows. Developers plugging into the Canva API can now offer users images where the text is actually legible in 55 different languages. More importantly, those users can instantly fix any mistakes without starting over.
This launch shifts the competitive moat for AI companies. Canva realized its massive library of user templates is a highly valuable, proprietary dataset that foundational models simply lack. By training on actual design files instead of raw internet photos, Canva built a specialized tool that beats generalist models at specific workflows. The company even had in-house designers guide the reinforcement learning process. Other platforms with massive, specialized user data will likely follow suit. They will train models on their own structured formats to bypass third-party API reliance, improve quality control, and cut massive compute bills.
For builders
Build tools with editable AI design layers
Developers can access Craft through the Canva public API right now using the model name craft_v1_2k. You can generate layered, editable designs instead of flat images for your users. Expect each API request to take roughly 50 to 60 seconds and cost AI credits.
Cut API costs with specialized in-house models
Canva built Craft partly to reduce its reliance on third-party models from OpenAI and Google. Founders paying massive API bills should note how Canva used distillation to create faster, cheaper internal models. Owning the model gives you total control over speed, quality, and profit margins.
Leverage proprietary data for specialized AI
Canva used its own library of millions of templates to train a model that understands layouts better than generalist photo models. Founders should look at their own user data. Structured, proprietary data is your best defense against massive AI labs.
My take
Canva built an AI to automate design by training it on the exact templates human designers made. It is a brilliant, ruthless move. I love that they realized flat pixels are useless for real work, so they weaponized their own proprietary data to build something OpenAI cannot easily copy. Designers trained the very system that will now do their heavy lifting.
Original reporting: TechCrunch. This is my rewrite and opinion.