Canva Debuts In-House AI Image Model Craft

Canva on Oct. 22 released Canva Craft, its first image generation model built and trained in-house by Canva Research, the company's AI lab. According to company officials, Craft generates images at 2K resolution across more than 30 aspect ratios and is available now inside Canva and to developers through a public API.

The company said Craft debuted as the fourth most advanced image generation model on DesignArena, a crowdsourced AI design evaluator, positioning Canva Research as the world's second most advanced AI lab for image generation by those rankings.

Craft outputs ranked No. 1 globally for People & Portraits and Landscapes & Nature, No. 2 for Abstract & Patterns and No. 4 for Products, the company said.

What Craft's DesignArena Ranking Measures

DesignArena ranks image models through crowdsourced head-to-head votes on AI design output. Canva reported that Craft debuted at No. 4 overall, and all three higher-ranked models come from OpenAI. That is the basis for Canva's claim to be the second most advanced AI lab for image generation.

The ranking is a launch-week snapshot. Arena leaderboards shift as votes accumulate and new models enter, so Craft's position will move. The category results say more about fit. Craft placed first among 87 models for People & Portraits and first among 85 for Landscapes & Nature, and Canva lists people, lifestyle scenes and landscapes among Craft's target use cases.

How Craft's 2-Stage Architecture Works

Most image generators produce pixels directly from a prompt. Craft splits the job in two. A planning model reads the prompt and writes a composition plan: which elements appear, where they sit relative to each other, their scale, their visual hierarchy and any text the image must contain. A separate rendering model then turns that plan into an image at up to 2K resolution across more than 30 aspect ratios.

The split targets text rendering and layout control, two areas where image models have historically struggled. Because Craft settles typography and placement before rendering, Canva said it produces legible text and more precise object placement. The plan also supports targeted edits, so a user can change one element without regenerating the whole image.

The same structure drives editability. Craft treats each image as a composition of layered elements and uses JSON bounding boxes, which Canva said makes conversion through Magic Layers faster and more accurate. Magic Layers splits a generated image into separate editable design elements. A Craft image can then be adjusted inside a Canva design the way a template can, while most generated images arrive as flat files.

Why Canva Trained Craft on Design Templates

Canva said most image models train primarily on photographs. Craft's training data leans on millions of human-made templates from Canva's content library, including posters, social posts, presentations, infographics, menus and other typography-heavy layouts. Templates carry information photographs don't: how layout, hierarchy and type fit together in a finished design.

The library gives Canva training material built around its own use case. General-purpose image models aim at every kind of image. Craft aims at the posters, ads and social graphics Canva users already make.

After pre-training, Canva refined Craft with supervised fine-tuning and reinforcement learning, using feedback from its professional designers on instruction following, typography and visual quality. Designers and researchers compared more than 25,000 image pairs against leading models in internal evaluations. Canva said that when automated scores and designer judgment conflicted, "we sided with the designer."

Why Craft Matters for Canva's AI Costs

Owning an image model gives Canva control over what each generated image costs to serve. Canva said more than a quarter of a billion people use the platform and create more than 400 designs every second. Every image routed through a third-party frontier model carries an inference bill Canva does not set.

Canva said model distillation cut the inference steps Craft needs to a fraction of the original model. A custom inference engine for the planning stage made Craft an order of magnitude faster than its first internal version. Those gains underpin Craft-fast and Craft-ultrafast, two upcoming variants Canva said will deliver high-quality images at a fraction of the cost of leading frontier models.

The launch follows Canva's August forecast cut, when AI serving costs pushed 2026 revenue growth guidance from 30% to 20%. Canva also said it had cut the cost per AI task by nearly 90% since April. Craft-fast and Craft-ultrafast have no release date or published pricing yet, so the savings remain Canva's projection.

How Craft Extends Canva's Agentic Marketing Push

Canva is selling Craft outside its own editor. Through the public API, developers can build Craft into marketing platforms, e-commerce sites, ad builders and print services. Canva also plans to connect Craft to AI assistants and automated workflows through the Canva MCP Server, so an agent can produce a launch graphic or social post as one step in a larger task.

That tracks Canva's 2026 acquisitions. Simtheory added agent management and Ortto added customer data and marketing automation. MangoAI and Doohly added ad optimization and digital out-of-home channels. With Craft open to agents, Canva can generate creative inside the same campaign workflows those tools run.

The API pitch has one gap. Products was Craft's weakest reported DesignArena category at No. 4, and e-commerce sites and ad builders are among the API customers Canva named.

Where Craft Still Falls Short

Canva named its own gaps. The company said future versions will improve fine print and technical drawings, which signals the first release's limits there. Reference-image input and more advanced editing are on the roadmap but not yet available, and the MCP connection is listed as coming soon.

Canva Craft Feature Breakdown

Canva said Craft uses a two-stage architecture that separates compositional planning from image rendering, aiming to give users greater precision and control.

Capability	Description
Two-stage architecture	Separates layout planning from rendering for compositional accuracy
2K resolution output	Generates images at up to 2K across 30+ aspect ratios
Magic Layers integration	Converts generated images into fully editable design elements
Multilingual prompts	Accepts prompts in any language; renders text in 55 languages
Design-trained model	Trained on millions of human-made templates emphasizing layout and typography

Recent Canva News

Canva has executed one of the most aggressive platform pivots in recent memory, completing five acquisitions in Q1 2026 and launching what it calls its largest product update since its 2013 founding.

The M&A spree included agentic AI platform Simtheory, customer data platform Ortto, creative analytics firm MagicBrief, ad optimization tool MangoAI, digital out-of-home platform Doohly and animation startup Cavalry.

COO Cliff Obrecht framed the Simtheory deal as accelerating Canva's evolution "from a design platform with AI tools to an AI platform with design and productivity tools at its core." The strategic intent: compete directly with established marketing automation and CDP vendors, backed by 265 million monthly active users and $4 billion in ARR.

On April 16, Canva launched Canva AI 2.0 at its annual Create conference, debuting a proprietary Canva Design Model capable of generating fully layered, editable output from a single prompt, along with conversational AI and integrations with Slack, Gmail and HubSpot.

The expansion, however, collided with cost realities. CEO Melanie Perkins slowed the AI rollout to rebuild architecture and reduce unit costs after demand outpaced expectations, forcing Canva to cut its revenue growth outlook by a third, to 20%.

AI Image Tools Reshape Design

Enterprise design teams are using AI image generation to restructure creative workflows from the ground up — but governance still lags behind the pace of adoption.

The underlying Canva Design Model previously powered Magic Layers, which has been used more than nine million times since its March 2026 beta, Canva said. AI products across the platform have been used more than 27 billion times overall, with usage tripling in the past year.

Governance Lags Behind Speed

Speed gains introduce governance risk. As earlier CMSWire reporting noted, enterprise creative teams are adopting AI tools faster than they can govern them. The result: brand drift, IP exposure and off-message content generated faster than review processes can catch.

Canva embeds brand standards into the generation layer itself. Other vendors — including Sitecore, Contentstack and Optimizely — have taken different approaches to maintaining brand alignment within generative workflows. For commercial-safe generation, Adobe Firefly remains notable for its licensed training data and reduced IP risk.

Canva Background

Founded in 2012, Canva provides an online design and visual communication platform positioned as an accessible alternative to traditional design software, serving individuals, marketers and organizations across education, government, nonprofit and commercial sectors.

Have a tip to share with our editorial team? Drop us a line: