news

Canva Craft debuts fourth on DesignArena, which Canva reports as second place among labs

Canva's first in-house image model launched October 22 with an Elo of 1340 on one crowdsourced leaderboard, three OpenAI models above it, and no independent evaluation attached.

By Daniel Ochoa · 24 October 2026

Canva's first in-house image model shipped on October 22 with a single external number behind every quality claim attached to it. Craft debuted fourth on DesignArena's image leaderboard at an Elo of 1340, and because the three models ranked above it are all OpenAI's, Canva converts that fourth place into a lab-level ranking: the company says Craft makes it the "world's second most advanced AI lab for image generation" on that board. Both statements are true at once. They just describe the same fourth-place result at two different levels of aggregation.

The model itself is available now inside Canva's Dream Lab and through the company's public API, generating at up to 2K resolution across more than 30 aspect ratios including custom dimensions. Canva Research, the in-house lab that built it, inherited staff from Leonardo.ai, the generation platform Canva acquired in 2024, and researchers from that team contributed to Craft.

A planner runs before the renderer, trained on templates rather than photographs

Craft splits generation into two stages. The first reads the prompt and produces a plan covering spatial relationships between elements, relative scale, visual hierarchy, and the text that needs to appear. The second renders the image from that plan. Canva says the training data was drawn from human-designed templates in its own library, posters, social graphics, presentations, infographics and menus, rather than from photographic corpora, which is a different starting point from most general-purpose image models. Outputs are meant to feed Canva's Magic Layers feature, which converts a generated image into separate editable design elements instead of a flat file. Two cheaper variants, Craftfast and Craft-ultrafast, are described as coming but are not out.

The category wins come with field sizes, except the one Craft did not win

Canva reports first place globally for People and Portraits out of 87 models, first for Landscapes and Nature out of 85, second for Abstract and Patterns, and fourth for Products. The field size is given for the two categories Craft topped and omitted for the Products placement, which makes that fourth-place figure harder to read than the others.

Crowdsourced Elo also moves. BetaNews reported Craft's DesignArena rating at 1330 on October 23, ten points below the 1340 Canva cited at launch the previous day, with the same three OpenAI models still above it. No confidence interval accompanies either figure in Canva's announcement, so the distance between fourth place and third is not quantified anywhere in the launch material. That is the recurring gap in image-arena launches, and it is the same issue we found when a rival arena separated its top two models by six Elo with intervals overlapping by fourteen points.

What is not in the announcement

There is no independent benchmark result, no prompt-adherence or text-rendering score from a third party, and no head-to-head sample set published alongside the ranking. The 2K ceiling is below the 4K output that Black Forest Labs claims for its current FLUX 3 image line, though the two are aimed at different jobs: Craft is built to hand back editable design layers inside a template editor, not to maximize resolution. Canva has also said it plans to expose Craft to third-party assistants through MCP, which is not live yet.

For a company that until now assembled its generation stack from other labs' APIs, including the Leonardo.Ai models it already owned, shipping a first-party model is the more durable part of this story than the leaderboard placement. The ranking will be re-sorted the next time OpenAI, Google or ByteDance ships.

Sources: Canva, TechCrunch, BetaNews.

Daniel Ochoa: Covers model launches, shutdowns and pricing changes as they happen. Reads deprecation notices for a living so you do not have to.