Canva Craft: Canva's First In-House AI Image Model Takes On Google, OpenAI, and Meta

Introduction: The Design Giant Builds Its Own Model

Canva, the design platform used by hundreds of millions of people, has quietly become an AI lab. On October 23, the company released Canva Craft — its first fully in-house image generation model, trained from scratch by its internal research division. Canva says Craft is competitive with comparable image models from Google, OpenAI, ByteDance, and Meta.

That's a bold claim from a company best known for templates and drag-and-drop editing rather than frontier AI research. But Craft arrives with receipts: a fourth-place overall finish on the DesignArena ranking — and first place in both the portraits and landscapes categories, per Canva.

For anyone choosing AI image tools in 2026, this launch matters. It signals that the biggest application-layer companies are no longer content to rent intelligence from OpenAI or Google — they're building their own, tuned to their users' actual workflows.

What Is Canva Craft?

Craft is a two-stage image generation model. Instead of generating pixels in one pass, it first plans the image — working out composition, proportions, and where text should sit — and only then renders the final picture. That planning step is why Craft handles the things generic image generators famously struggle with: layouts with multiple elements, posters with headlines, and designs where text placement is as important as the imagery.

The headline specs:

• Resolution up to 2K — high enough for print-ready assets, not just social media thumbnails.
• More than 30 aspect ratios — from square posts to widescreen banners, natively.
• Text rendering in 55 languages — addressing the garbled-text problem that still plagues many image models.
• Magic Layers integration — generated images can be broken into editable layers inside Canva, so you can tweak a headline or swap a background without regenerating everything.

The Template Advantage: Trained on Design, Not Photos

The most interesting technical decision is the training data. Rather than leaning on the usual web-scale photo corpora, Canva Research trained Craft primarily on millions of design templates — the structured, layout-aware artifacts that are Canva's actual core asset.

This changes what the model is good at. Models trained mostly on photographs excel at photorealistic scenes. Craft's template-heavy diet means it internalized design structure: grid systems, typography hierarchies, whitespace, and how a poster or presentation slide is actually organized. For the work Canva's users do — social graphics, marketing materials, decks — that's a better prior than knowing what a sunset looks like.

The project also drew on talent from Leonardo.ai, the AI image generation startup Canva acquired in 2024. Leonardo's researchers contributed directly to Craft's development, making this the first visible fruit of that acquisition beyond Leonardo's own product line.

How It Performs: DesignArena Results

Canva's credibility check comes from DesignArena, an independent platform that pits image models against each other on design-oriented tasks. According to Canva, Craft placed:

Treat vendor-cited rankings with the usual caution — Canva is choosing which numbers to highlight. But a fourth-place overall finish for a first-generation in-house model, built by a company that isn't a frontier lab, would have been unthinkable a year ago. The gap between "frontier model" and "good enough, integrated, and cheap" keeps closing.

Where to Get It: Dream Lab and the Public API

Craft is rolling out in two places:

• Dream Lab (canva.com/dream-lab) — the consumer-facing generation tab in Canva's editor, where Craft sits alongside the third-party models Canva previously relied on.
• The Canva API (canva.com/developers) — developers can now generate 2K images in 30+ aspect ratios programmatically and receive layered, Canva-editable output. This is the piece that turns Craft from a feature into infrastructure: apps can generate designs that users can continue editing in Canva via Magic Layers.

Also on the roadmap: MCP integration, which would let chatbots like ChatGPT and Gemini generate Canva designs directly through the Model Context Protocol. When that ships, "make me an Instagram post about this" in any MCP-capable assistant could route straight into Craft.

The business logic is straightforward. Running your own model is cheaper at Canva's scale than paying per-generation fees to a frontier lab, and it gives the company control over quality, latency, and roadmap. It's the same calculation Adobe made with Firefly — own the model, own the cost curve, own the differentiation.

Why It Matters for the AI Tools Market

Craft is the latest and clearest sign that the AI tools market is bifurcating. On one side: frontier general models from OpenAI, Google, Anthropic, and Meta, competing on raw capability. On the other: application-native models built by the platforms themselves — Adobe's Firefly, Canva's Craft — competing on workflow fit, price, and integration depth.

For most people choosing an AI image tool, the second category is increasingly the pragmatic answer. A model that hands you a fully layered, editable 2K design in the exact aspect ratio you need — inside the editor you already use — beats a marginally better raw model that outputs a flat PNG you have to reconstruct by hand.

Watch for the same pattern to spread: expect every major SaaS platform with a visual product to ship an in-house generation model in the next 12–18 months. The question for tool shoppers is shifting from "which model is smartest?" to "which model lives closest to my work?"

Frequently Asked Questions