Canva Launches First In-House AI Image Model, Challenging the Giants

Canva unveiled Canva Craft on Wednesday, October 22, 2026, its first image generation model built and trained entirely by its internal research team. The model's debut positions the company as the world's second most advanced AI laboratory for image generation, trailing only OpenAI, according to independent evaluation by DesignArena, a crowdsourced platform for ranking visual AI models.

What makes Craft distinctive is not just image quality—though early benchmark results placed it fourth overall on the DesignArena leaderboard, ahead of models from Google, Meta, xAI, and ByteDance—but its architecture. While most AI image generators (DALL-E, Midjourney, Imagen) map a text prompt directly to pixels, Craft splits the process into two stages: first, it plans the composition; then, it renders the image.

The two-stage architecture

In the first stage, Craft interprets the user's prompt and produces a detailed "composition plan," specifying which visual elements should appear, their spatial relationships, relative scale, visual hierarchy, and information about text to be included. This plan functions as a semantic skeleton before rendering begins.

In the second stage, a dedicated rendering model translates that plan into a high-fidelity image at up to 2K resolution, in more than 30 aspect ratios—spanning social media posts to corporate presentations. According to the company, this approach enables the generation of images with legible text, precise object placement, and consistent visual relationships, something that has been historically challenging for general-purpose image generators.

The key difference is analogous to what separates a painter working by trial and error from an architect who first draws the blueprint: Craft thinks before it draws. This is not merely an aesthetic distinction—it is a functional advantage for a product like Canva, whose users want to generate a banner, edit it, change colors and text without recreating the entire image.

Training on templates, not photographs

Another technical differentiator lies in the training data. While most image generation models have been trained predominantly on photographs, Craft learned from millions of human-designed templates from Canva's own library—posters, social posts, presentations, infographics, menus, and typography-rich layouts. The idea, according to the research team, is to teach the model not just what images look like, but how visual elements organize into functional layouts.

Canva evaluated the model internally with its professional designers, who compared 25,000 pairs of images in blind evaluations against leading market models. When automated assessments and human judgments diverged, the company chose to prioritize the designers' judgment.

In DesignArena subcategories, Craft ranked number one globally for "People & Portraits" (among 87 models) and "Landscapes & Nature" (among 85 models), second for "Abstract & Patterns" and fourth for "Products".

The economics behind the model

For Canva, developing its own model is also an economic imperative. The company operates with 240 million monthly active users and roughly US$ 3 billion in annual revenue—but running third-party AI has proven enormously expensive. In May 2026, Canva cut its 2026 growth projection from 30% to 20%, and its valuation fell from US$ 42 billion to US$ 35 billion, according to external analysts, in part due to unsustainable costs of relying on frontier models from OpenAI, Google, and other companies.

With Craft, the company regains control over cost, quality, and performance. It is also developing two cost-efficient variants—Craft-fast and Craft-ultrafast—that promise competitive quality at a fraction of the cost of frontier models, thanks to advanced model distillation techniques and a custom inference engine optimized for the planning stage.

Integrations and API

Craft is already available to all Canva platform users through the Dream Lab tab and will be offered via API for developers. The company also plans to integrate the model into third-party chatbots, including ChatGPT and Google Gemini, via the Model Context Protocol (MCP). Images generated by Craft can be automatically converted into layer-editable designs using the Magic Layers tool, launched earlier in 2026.

Canva's strategy is ambitious: instead of competing on photorealism, it competes on practical utility. In commercial design, the user does not simply want a beautiful image—they want a layout that can be revised, adjusted, and shipped. If Craft can deliver that at 400 million designs per second across the platform, Canva's real bet is that it is not just launching an image model, but redefining what an image generator should be.

Sources: The Verge, TechCrunch, Canva Newsroom

✓ Independent sources cross-checked and verified before publishing