Canva launches Craft, its first in-house image model

By VK

Published 26 Oct 2026, 12:22 UTC

Reporting from TechCrunch, The Verge

Canva says Craft plans an image's layout before rendering it, was trained heavily on its templates, and is open to all users and via its public API.

What it means for founders

A new API option for design-heavy output. If your product makes ads, social posts, menus or print assets, benchmark Craft against Google's Nano Banana 2.1 and OpenAI on your own prompts, especially where legible text matters. Canva pitches its API for that kind of side by side test.
Layers may be the bigger draw. An image that splits into editable parts could save a cleanup pass. If your workflow ends in Canva anyway, generating there may beat importing flat files from Adobe Firefly or elsewhere.
Watch the cost curve. A design platform building cheaper in-house tiers suggests image generation prices may keep falling. Short commitments to any single image provider look safer while that continues.
Signals to track: the release dates of Craft-fast and Craft-ultrafast, the arrival of Craft in Canva's MCP server, and whether Craft holds its Design Arena rank once rivals ship updates.

The story

Canva released Canva Craft on Thursday, October 22, the first image generation model the design company has built and trained itself. It is live for every Canva user worldwide through the Dream Lab panel, and developers can call it through Canva's public API, according to the company's announcement. The launch puts Canva in direct competition with image models from OpenAI and Google.

How Craft works

Canva says Craft splits generation into two steps. First, a planning model maps the prompt into a layout: what goes in the frame, where, at what size, what leads the eye and any wording. A second model then draws the image from that plan at up to 2K resolution. Canva lists more than 30 preset aspect ratios plus custom sizes, and says Craft can render text in 55 languages.

The company argues that planning first is what lets Craft produce readable lettering and place objects precisely, and lets users change one element without disturbing the rest. Because outputs are structured this way, Canva says they convert cleanly into editable layers through its Magic Layers feature.

Trained on design templates

Canva says rival models lean mostly on photographs, whereas Craft's training gave more weight to design: millions of human-designed templates in its library, from posters and slide decks to menus and infographics. It adds that feedback from its own designers guided later fine-tuning and reinforcement learning. Researchers from Leonardo.ai, which Canva bought in 2024, worked on the model, TechCrunch reported.

Canva says Craft entered the crowdsourced Design Arena leaderboard in fourth place at launch, and The Verge reported that all three models above it came from OpenAI.

Why Canva wants its own model

TechCrunch noted Canva has used several image models and suggested owning one would likely trim API bills; Canva, it reported, also cited tighter quality and speed control. Canva says two lighter versions, Craft-fast and Craft-ultrafast, are coming and will generate images far more cheaply than leading frontier models. It also plans to expose Craft to AI assistants through its MCP server, a link it labels as coming soon.

TechCrunch also noted that Canva was heading for $4 billion of revenue in 2025 and has since cut its 2026 growth forecast, a move that outside analysts say lowered its valuation to $35 billion from $42 billion.

What we don't know yet

What Craft costs per image through the API, and how much cheaper the fast variants will be.
When Craft reaches the MCP server, and which assistants it will work with first.
How Craft handles fine print and technical drawings, which Canva names as areas it is still improving.