AINews
Canva trains its own image model to lay out a design before drawing it
Canva Craft plans scale, layout and text first and renders second; the company's quality claims have not been independently tested.
What the reporting says
TechCrunch reports that Canva has released Canva Craft, an image model built by its own research team. According to the report, the model works in two steps: it first plans the layout, relative scale and text of a design, then renders the image at up to 2K resolution. The Verge adds that the model is available now to all Canva users worldwide, that it ranks fourth on a crowdsourced image leaderboard behind three OpenAI models, and that API access is planned. Both outlets present Canva's claims of better precision and cleaner edits as the company's own, and neither reports independent testing of them.
Read the original reporting
TechCrunch
Canva releases its first image generation modelMy Read
Commentary: opinion, not reporting.
Planning the layout before rendering is a sensible answer to the two things image models get wrong most often: legible text and where objects sit. It is the same logic as writing the design before the code, and it suits a product whose customers want to edit what they get, not just admire it.
A fourth-place finish on a crowdsourced leaderboard is respectable, but I would want to know how much of it rests on design prompts and how much on general image prompts. The precision and clean-edit claims are Canva's own, so I would treat them as a hypothesis until someone tries to break the model on purpose.
TechCrunch suggests a cost motive, since a company that has relied on several outside models could lower its bills with one of its own. That is a reasonable guess about incentives, not something Canva has said. What the reports leave out is what the moderation and reporting controls cover.
Updates
Shortened and rewritten in the site's current format: a brief summary of the reporting with a link to the original, then commentary.
- The longer summary of the reporting was replaced by a short one that names the outlet and links to it.
- The commentary now appears under its own heading, My Take.
- No facts were added.
Discussion
Is a model that plans a layout before it draws a better bet for design work than one that simply generates?
Covered by 2 of 10 outlets we read, on Techmeme, 17 points on Hacker News.