Canva has released its first in-house AI image generation model, Canva Craft, and it goes straight up against image models from OpenAI, Microsoft, and Google. The Verge AI reports that the model uses a two-stage architecture. It plans a design first and only then generates the image. Canva says this approach gives users “greater precision, consistency and control.”
Until now, Canva mostly built AI features around other companies’ models. Owning the model is a different strategy. Canva now controls how it’s trained, how it’s tuned, and what it costs to run.
🧭 What Canva Craft Brings to the Table
Here’s what matters most in the launch, based on The Verge AI’s coverage:
- Plan first, render second. The two-stage system decides the layout before any pixels exist. Canva says this makes Craft more accurate on complex images that need readable text and exact object placement. Those are two areas where image generators still often struggle.
- Up to 2K resolution and more than 30 aspect ratios. That covers most formats marketers deal with every day, from vertical stories to wide banners, without a lot of cropping.
- More than marketing assets. Social posts and marketing content are the obvious use cases. Canva says Craft can also produce realistic scenes, digital art, and environmental imagery.
- Trained on real templates. Canva says it trained Craft on “millions of human-made templates” from its content library, including social media posts, presentations, infographics, and menus. This is Canva’s clearest advantage. Few AI labs have that much structured, design-focused training data.
- Built-in safety tools. Canva says the model includes “robust input and output moderation, privacy and safety controls, and reporting tools.” Business customers will expect exactly that before they put it into brand workflows.
- Agent access through MCP. Canva plans to let AI agents connect to Craft through the open Model Context Protocol, a standard way for AI assistants to call outside tools. In practice, an assistant could hand off a design job to Craft on its own.
📊 How It Stacks Up
Craft currently ranks fourth on Design Arena’s leaderboard, a crowdsourced benchmark where people vote on AI-generated designs. The three models above it all come from OpenAI: GPT-Image-2.5 Sunburst, GPT-Image-2.5 Flare, and GPT-Image-2.
Fourth place for a first model is a strong debut. It also means OpenAI still sets the pace in this category. Canva’s pitch isn’t really “best raw image quality.” It’s “the best fit inside a design workflow,” and that’s a smarter fight for Canva to pick.
🛠️ Part of a Bigger AI Push
Craft builds on a run of recent AI updates. In April, Canva added prompt-based editing, which lets you create or change a design just by describing what you want. It also added layers to AI-generated images, so you can edit one part of an image and leave the rest alone.
Put those together and the strategy is clear. Canva wants AI output that stays editable, structured, and on-brand. A one-shot image you can’t adjust isn’t the goal. A model that plans layouts before it renders them fits that approach well.
🌍 Availability and What’s Next
Canva Craft is available now to all users globally. To try it, open the three-dot option in Canva’s left-hand menu and select Dream Lab.
Canva also plans faster, more efficient versions called Craft-fast and Craft-ultrafast. It hasn’t announced dates for those or for the MCP integration.
⚠️ Worth Keeping in Mind
- No pricing breakdown. The announcement doesn’t say whether usage limits differ between free and paid plans.
- Benchmarks only tell part of the story. Design Arena is crowdsourced and shows preferences, not reliability on specific jobs like legible text.
- Training data questions. Training on “human-made templates” raises the obvious question of how the creators behind those templates are credited or paid. The announcement doesn’t address it.
🔭 Why It Matters
Canva sits where hundreds of millions of non-designers already make their content. A model built for that audience, trained on that kind of content, and connected to AI agents could take real share from general-purpose image generators in everyday business design work. The next thing to watch is whether the fast variants and MCP access arrive soon enough to keep up with OpenAI’s release pace.
The full announcement details are available in The Verge AI’s original report.