Futureweb

Artificial Intelligence

Canva shipped Craft, its first in-house image model. The ranking it cites is DesignArena's, and the training claim is the company's.

Canva says Craft plans a layout, then renders up to 2K in more than 30 aspect ratios. The DesignArena rank is Canva's claim. TechCrunch reports the launch.

The short version

Canva says Craft is its first image model, available in Dream Lab and on its API. It plans a composition, then renders at up to 2K in more than 30 aspect ratios. Canva says that ranked its lab second on DesignArena. TechCrunch and The Verge report the launch. This page did not rerun the board.

A company model, with a board attached

Canva's newsroom post says it is unveiling Craft, developed and trained by Canva Research, built for design, and available now inside Canva and to developers. It says images come out at 2K in any aspect ratio the product supports, which it puts at more than 30. It says Craft debuted fourth on DesignArena, which it calls the largest crowdsourced assessor of AI design, and that the lab is therefore second in the world at image generation. TechCrunch's October 22 story and The Verge's October 22 story both lead with the two-stage architecture and the same resolution and ratio figures. They are accounts of the launch. They are not a second laboratory's score.Phonon is a different model, with a weight file.

Plan, then pixels

Canva says a conventional generator goes from prompt to pixels, and Craft does not. A planning model writes the composition, including spatial relationships, scale, hierarchy, and the text that should appear. A rendering model turns that plan into the image. The stated reason is control: change one element without redrawing the rest, and keep type legible. TechCrunch adds that the outputs can be turned into layered designs through Canva's existing image-to-design tool. The Verge says the same planning step is the company's explanation for precision. A reader who wants a bake-off will not find one in these three pages. A reader who wants the product claim will.a reranker scores a list.

What the training claim is

TechCrunch says Canva told it the model was trained on designs from Canva's library, including posters, social posts, presentations, infographics, menus, and type-heavy layouts, rather than on photographs. That is a claim about the corpus. It is also a claim about why the model should be better at design than a photo model. This page cannot see the corpus. It can see that the claim is attributed to the company and printed by TechCrunch. The newsroom post itself stresses editable output and design understanding. The photograph contrast is the interview's, carried by TechCrunch.what a training cutoff does not mean.

The leaderboard is not a unit

DesignArena, on Canva's description, uses people to judge images. Fourth place for the model and second place for the lab are only coherent together if the models above Craft belong to one other organization, which is what Canva says. A different cut of the same board, or a different week, can move both numbers. The Verge does not rest its story on the rank. TechCrunch names the competitors Canva says it is chasing, as a category, without adopting the rank as TechCrunch's measurement. This page treats the rank as a sentence Canva is willing to have quoted, and it says who measured.an evaluation a second lab can run.

API now, assistant hook later

Canva says the public API is how other products can call Craft, and that an MCP connection so assistants can ask for an image is coming soon, not shipping in the same sentence. Coming soon is a schedule. The API is the availability claim a developer can test. This page does not print the request. It also does not list price. A model that is on an API is not thereby the default model inside every Canva button. The newsroom says Dream Lab is the place a person starts.Open Responses is a spec, not this model.

What the outlets do not add

Neither TechCrunch nor The Verge reports a downloaded weight file, a license for running Craft on your own hardware, or a parameter count. Phonon, filed yesterday, was that kind of release. Craft is a service Canva hosts. Comparing them as if both were a file you can vendor would invent a distribution Canva has not announced. The Verge's frame is a company entering a market that already has models from larger labs. That frame can be right and still leave the weights unpublished.on-device inference keeps a different line.

Who is in scope

Canva users, as the people who can open Dream Lab. Developers integrating the API, as the people who can call it. Readers of the DesignArena claim, as the people who need the rank attributed. People who wanted a local weight file are out of scope, because no such file is in these sources. Competitors on the board are in scope only as the context Canva cites.

What stays on the record

Canva newsroom: Craft, first in-house image model, Dream Lab, public API, MCP later, 2K, more than 30 aspect ratios, plan-then-render, DesignArena fourth for the model and second for the lab, on Canva's account. TechCrunch, October 22: two-stage architecture, training on design templates rather than photographs, layered output via Canva's existing tool. The Verge, October 22: same architecture and availability, no independent score. No weight file. No request printed here.

What would move the row

A DesignArena page that shows the week of the ranking, or a Canva note that publishes weights, would change what kind of object Craft is. An outside test of the text-rendering claim would be the measurement these three stories do not contain. Until then the record is the newsroom post and two launch accounts. The short answer this page will stand on is this: Canva says Craft is its first image model, available in Dream Lab and on its API. It plans a composition, then renders at up to 2K in more than 30 aspect ratios. Canva says that ranked its lab second on DesignArena. TechCrunch and The Verge report the launch. This page did not rerun the board.NASK is a research deal, not this model.

Sources

The reports this brief is filing. Futureweb did not republish them.

  1. Canva
  2. TechCrunch
  3. The Verge

Questions

Is Craft the same thing as a photo model Canva resold?

Canva says Craft is the first foundation model its own research group built and trained from scratch. TechCrunch says the training material was human-designed templates from Canva's library, such as posters and slides, rather than a photograph corpus. That training claim is the company's, via TechCrunch.

Where can a person use it?

Canva says every Canva user, in Dream Lab, and developers via the public API. It also says a connection for assistants through its MCP server is coming later, not part of today's switch. The Verge reports the same availability story at the level of the product, not the API fields.

What does two-stage mean, in Canva's words?

The first stage writes a plan: which elements, where, how large, what text. The second stage renders that plan. Canva says the split is why text and placement should hold still when one element changes. The Verge and TechCrunch both attribute that architecture to Canva. Neither published an independent test in the stories this page used.

Who is second, and at what?

Canva says the model debuted fourth on DesignArena and that this makes Canva Research the second most advanced lab for image generation, because the three models ahead of it are from one other lab. A leaderboard place is the board's result under the board's rules. It is not a measurement this paper repeated.

Does this page include the API call?

No. A model name on an endpoint is a developer detail. The fact for this page is that Canva says the model is on the public API, not the shape of the request.