Canva Craft Ranks Fourth on DesignArena but Trails in Marketing

By James Calloway

Canva Craft, Canva's first proprietary AI image model (one it develops and operates itself), ranks fourth overall on DesignArena, trailing only OpenAI's ChatGPT Images in early testing. Canva says the model is now available globally to all Canva users and built specifically for designers.

That launch gives Canva an in-house model designed around layout rather than another generic image generator added to its platform. It matters because Canva, a beginner-friendly design program, has incorporated AI tools and third-party AI models for years. Its first creative AI feature, Magic Media, debuted in 2023.

The new release extends that effort into model ownership, training data, and editable output. CNET's report did not provide a launch date, pricing, model size, or usage limits, so the global availability claim says little about cost or performance at scale.

Why did Canva build its own model?

Stefano Corazza, Canva's head of AI research, told CNET that two decisions shaped Craft: Canva's acquisition of Leonardo.Ai and an internal increase in AI research. He said both moves helped the company build its own AI image model, one made specifically for designers.

Corazza also presented in-house development as an efficiency measure. Keeping the technology inside Canva can reduce cloud costs and shorten inference, meaning the wait between submitting a request and receiving an image, he said. The source material provides no figures for either benefit and no independent speed or savings test.

Corazza described planning as central to the approach: “[The AI] can actually create a whole plan for how the image is going to look like, all the layout. If you think about designers, that’s exactly what they do right when they create a design.”

He compared the model with designers who carefully place elements to create a story. He said Craft does basically the same thing and mimics the design process.

In practical terms, Canva now controls a model built around its own workflow. Without measurements, Corazza's efficiency argument remains a claim rather than a demonstrated result.

How does the two-stage model work?

Canva said in a blog post that Craft uses a two-stage architecture, meaning it develops a plan before producing the final design. Combined with Canva's prompt-writing tools, the model can “reason” by working out composition and layout before generation. That sequence gives the system an intermediate planning step before final rendering.

Corazza said Craft trains on three disclosed categories of material:

• Canva's own library.
• Licensed and public databases.
• Data from Canva users who opted in, meaning they agreed to allow their contributions to train the model.

Canva presents the combination as a route to designs that are more complex and comprehensive. Corazza used “ideally” for that result, and the materials cited by CNET contain no side-by-side examples, testing protocol, or measurements of complexity.

The recipe matters because artists have raised concerns about training data and copyright protection. Canva's disclosure identifies broad categories rather than named datasets, volume, licensing terms, or safeguards for checking conflicting rights. The report records Canva's account of its sources, not independent verification that its use of them is legally or artistically uncontroversial.

The architecture prioritizes deliberate planning and editable structure, but the evidence remains mostly Canva's own characterization. Independent reviewers still need comparable prompts, output examples, and reproducible performance data.

How strong are the early benchmark results?

DesignArena, an image-model comparison benchmark, places Craft fourth overall, behind only OpenAI's ChatGPT Images. In that ranking, Canva Craft also beats Google's Nano Banana and Microsoft's MAI-Image models.

The platform's broader results split by use case:

• Portraits: Canva Craft ranks highly, though CNET gave no exact position.
• Product-focused imagery: Canva Craft also ranks highly.
• Marketing: Canva Craft trails the field in the category where many people use Canva.

Those results strengthen the early case for a designer-focused model, but the source gives no scores, ranking date, test prompts, or judging method. Without those details, the position establishes a relative result rather than a reproducible comparison.

The marketing gap carries particular weight because that category overlaps with a large Canva user base. The early ranking looks strong overall, while its reported marketing gap sits close to Canva's core market.

Why does editing matter?

CNET framed editability as a response to a common generator failure: even a carefully written prompt often produces an image or video that is almost, but not quite, right. Canva has previously offered Magic Layers, which separates an AI design into its basic elements and lets users adjust them.

Canva describes Magic Layers as a technologically complex solution to a major problem for AI enthusiasts. Canva Craft builds on that foundation and allows users to edit its generated results. Editing generated media matters because a strong first result does not remove the need for adjustments.

The report does not specify which elements users can change in Craft, how granular those controls become, or how the editing workflow compares with Magic Layers. Those omissions limit any assessment of Craft's practical control over its output.

Editability may be Craft's most useful advantage, but the report leaves the control set undefined. A strong benchmark position cannot compensate for an unclear revision process.

What does the creator controversy change?

CNET described creative AI as controversial, citing years of concerns from photographers, illustrators, and artists. Their complaints include alleged training on copyrighted work without permission, the creation of abusive material, and arguments that AI has been offered to executives as a replacement for real artists.

Adobe and Canva have maintained that AI is simply a tool they want to offer their creators. That position does not resolve the disputes surrounding training, substitution, or provenance, meaning the evidence that an image came from a particular source.

Eric Wittman, CEO of the popular photo editing and sharing platform VSCO, replied to Canva CEO Melanie Perkins' LinkedIn post about Canva Craft. He said: “real photographers, not AI, will always rank No. 1 for portraits.” The remark turned Craft's portrait benchmark strength into a dispute over whether quality leadership should carry any weight.

CNET also reports that photographers face AI problems on every design platform, including VSCO. As models improve, creators must consistently verify that their work is human-made rather than AI-generated. Digital verification methods, systems for confirming a media file's origin, remain sporadic, leaving creators to navigate a more complicated web.

Craft's opt-in policy addresses consent for the Canva user data used in training. It does not settle broader questions about copyright, abusive outputs, professional substitution, or proof of human authorship. Those are platform-level problems that a portrait benchmark cannot answer.

Canva's next test is practical: whether Craft can win marketing work through editable output while creators and platforms establish more reliable human-made verification.