"It looks incredible until you try to read it." A designer I work with said that a few weeks ago, zoomed all the way into an AI-generated poster where the big confident headline spelled a word that has never existed in any language. We both laughed. Then we both quietly closed the tab and opened Figma.
That has been the running joke about image models for three years. They could render a photo-real anything, a golden retriever astronaut, a cathedral made of coral, but ask them to lay out a sentence and decide what your eye should land on first, and they fell apart. The easy read on Canva's newest release is that it's one more model joining a crowded race. I think that read misses the only part that matters.
What Canva actually shipped
On October 22, 2026, Canva launched Craft, its first in-house image generation model. The detail worth your attention is how it works. Per Canva's announcement, Craft runs in two stages: a planning model first drafts the composition, the object placement, scale, visual hierarchy, and the text, and only then does a rendering model turn that plan into the final image. It was trained on millions of human-made Canva templates (menus, posters, presentations) rather than mostly on photographs, outputs up to 2K resolution, and, as reported by Digital Trends the same day, debuted fourth on the DesignArena leaderboard, which Canva says makes its research lab the second most advanced for image generation as of October 2026.
Leaderboard positions churn. By the time you read this, some lab will have leapfrogged it. That's not the story.
The architecture is the story, not the ranking
Here's what I think actually happened, and it's an opinion, not a benchmark: a major design company decided the way to make a better design image is to make the model think about structure before it thinks about pixels. That is how designers work. We don't open a file and start with color. We start with hierarchy. What is this screen, what do you look at first, where does the eye go next, what can wait. The picture is the last thing that happens, not the first.
Every pixel-first model before this treated layout as an accident that emerged from denoising noise, which is exactly why the text came out as alien runes and three buttons fought for the same corner. A planning pass that reasons about hierarchy and placement before it renders is the first image release in a while that I'd file under thoughtful rather than gimmick. Canva didn't just train a prettier renderer. They put a designer's order of operations into the pipeline.
Why this matters if you design products, not posters
The limits are real. An image model doesn't ship your product. No product designer hands off a PNG and calls it a feature. Craft is aimed at Canva's poster-and-presentation crowd, not at your component library.
But a model that plans structure before it renders is a preview of the tools coming for the rest of us. This is the same bet underneath generative UI: that the valuable move is composing a correct structure, and the surface is almost a side effect. When I designed the loan screens for the Monety fintech app, the hard part was never making it look good. It was deciding what a stressed borrower in a hurry should see first, and what to hide until later. Hierarchy was the entire job. Color was fifteen minutes at the end.
So here's the uncomfortable part. As the render layer gets automated, the human value concentrates in the plan. The part of our work these models are still worst at is the part Craft is quietly admitting is the hardest to get right: deciding what matters. That judgment is the taste AI can't automate, and it's about to become the whole job description.
If you want one habit out of this: stop grading these tools on whether the picture is pretty. Grade them on whether they understand structure. And get fluent at describing the hierarchy you want in words, because "tell me what should be seen first" is turning into the real interface to design AI.
The models are finally learning that a design is a decision before it's a picture. The open question is whether we're ready to be paid for the decision instead of the picture.
Share on
