Canva Craft (Canva's first AI image model) Review

The video review. Watch it on YouTube

What Canva Craft does

Canva Craft is the first foundation model that Canva Research trained from scratch. Before its October 22, 2026 launch, Canva's image features ran largely on models from outside providers. Craft changes that: Canva now owns the full pipeline from prompt to pixel.

The model works in two stages. Stage one, called the image planning model, reads your text prompt and produces a compositional plan: where objects sit, their relative scale, the visual hierarchy, and which text appears and where. Stage two, a dedicated rendering model, takes that plan and paints the finished image at up to 2K resolution. The internal identifier for the pair is CRAFT-1 2K.

Canva built it for design work. The company's own framing is posters, social graphics, slide decks, restaurant menus, infographics, and other type-heavy commercial layouts. It is explicitly not aimed at photographic realism or contest-winning art.

Key features

Craft renders text in 55 languages and accepts prompts in any language, which matters if you are making a menu or a local ad outside English. It supports more than 30 aspect ratios or a custom ratio, and the maximum output resolution is 2K.

The training data is a real differentiator. Craft was trained on millions of human-made design templates from Canva's library, then fine-tuned with supervised learning and reinforcement learning using feedback from Canva's own professional designers. Canva also reports 25,000-plus head-to-head image pairs judged by its designers against leading models, though those comparisons have not been published in detail.

On the integration side, Craft output feeds directly into Canva's Magic Layers tool, which converts a flat generated image into individually editable elements: text, colour, layout. Researchers from Leonardo.ai, which Canva acquired in 2024, contributed to the model, and the older Phoenix model now sits alongside Craft in Canva's internal mix.

Distilled variants called Craft-fast and Craft-ultrafast are described as promised soon or in development in the launch coverage. Neither was available at the time of this review.

How it sits in your Canva workflow

You access Craft through Dream Lab inside Canva's web app. The navigation path is the homepage, then More, then Dream Lab. According to the launch coverage, it is available to all Canva users, including those on the free tier, with no separate purchase.

The Canva site also lists desktop downloads for Windows and Mac, so the broader editor is available beyond the browser. For developers, Craft is reachable through the Canva API, though the API is still in preview as of the launch-day reporting.

A Model Context Protocol server connection is described as coming, which would let AI assistants call Craft as part of larger agent workflows. That integration is not live yet.

How it compares to Midjourney and Adobe Firefly

On DesignArena, Craft debuted at fourth overall with an ELO score of 1,341. That places it above models from Google, Microsoft, Meta, xAI, and ByteDance. Three OpenAI models sit above it. The rankings are preliminary and shift as votes accumulate.

Midjourney 7.0, at $10 a month, gives you a four-option grid and is aimed at artistic and photographic generation. Craft's pitch is different: spatial control, legible text, and commercial layout. If your brief is a flyer for a bakery opening, Craft's design-template training is the more relevant fit.

Adobe Firefly 5 is the other major comparison. It is commercial-safe and embedded in Photoshop and Illustrator. Craft is embedded in Canva's template ecosystem, and the handoff from generation to editable design is tighter because you stay in one editor.

What is missing and what to watch

No technical report, parameter counts, training-compute figures, or architecture white paper was released with the launch. Canva's speed and cost claims, including that a custom planning-stage engine made it an order of magnitude faster than its first internal version, are self-reported and have not been independently verified.

One practical caveat from an independent write-up: text drawn into a generated image is still part of the image. It does not automatically become an editable text box when a client asks you to change three words. You need to run Magic Layers as a separate step, then inspect the result to confirm it separated cleanly.

The API is still in preview, and full general-availability timing is not specified. Canva shelved some heavily promoted AI products in August 2026 over token costs and cut its 2026 revenue-growth forecast from 30 to 20 percent because third-party inference costs were unsustainable at its scale of 265 million monthly users. If you are building a production pipeline, the preview status is a real risk.

Pros and cons

Canva Craft (Canva's first AI image model) alternatives

• Midjourney 7.0: Aimed at artistic and photographic generation with a four-option grid at $10 a month. Craft targets commercial layout, legible text, and stays inside Canva's editor.
• Adobe Firefly 5: Commercial-safe and embedded in Photoshop and Illustrator. Craft lives in Canva's template ecosystem, so the generation-to-edit handoff is tighter if you already work in Canva.
• Leonardo.ai Phoenix: Now part of Canva's internal model mix following the 2024 acquisition. Phoenix coexists alongside Craft rather than competing with it.

FAQ

How this review was made: researched from Canva Craft (Canva's first AI image model)'s public website, docs and independent sources, and written with the help of AI. We don't claim hands-on testing. How we review