Canva Craft shipped on 22 October 2026: Canva’s first image generation model built by its own research team, available to all users through the Dream Lab tab and through Canva’s public API.
Two things about it are worth a designer’s attention. One is how it handles text and layout, which is a real technical difference rather than a marketing line. The other is that Canva can answer the question most image-model vendors deflect.
It plans the image before it draws it
Canva describes a two-stage pipeline. In the first stage the model reads the prompt and produces a plan: spatial relationships between elements, relative scale, visual hierarchy, and the information needed for text generation. Only then does the second stage render — up to 2K resolution, in a possible 30 aspect ratios.
That ordering is the whole point. Most diffusion models generate text as texture that happens to look like letters, which is why they produce beautiful posters saying GRAND OPENNIG. Deciding “this is a headline, it sits here, it reads this” before committing pixels is a different problem, and a tractable one. Canva says the approach yields accurate object detail and legible text, and that users can make precise edits without disturbing the rest of the image. Canva also claims text rendering in 55 languages.
The training corpus follows from that: rather than photographs, Canva used human-designed templates from its own library — posters, social posts, presentations, infographics, menus and typography-rich layouts. A model trained on designs, to make designs.
There is a downstream consequence. Because Craft’s output is composed rather than merely rendered, it feeds Canva’s Magic Layers — the image-to-design transformer Canva launched in March 2026, which separates a flat image into movable elements while preserving the relationships between them. Generate, then take it apart and edit it like a real document. That round trip is the thing Canva has that a raw model endpoint does not.
Researchers from Leonardo.ai, which Canva acquired in 2024, contributed to the model. Two cheaper variants, Craft-fast and Craft-ultrafast, are described as coming soon, along with image generation inside third-party chatbots such as ChatGPT and Gemini via Canva’s MCP server.
The benchmark claim, and what we could not check
Canva says Craft debuted fourth overall on DesignArena, and first in portraits and landscapes. Press coverage reports its Elo slipping from 1340 at launch to 1330 by 23 October.
We could not confirm this independently. designarena.ai renders its leaderboard client-side, so a fetch returns the page shell with the chart containers empty and no model names or scores in the markup. The ranking is Canva’s claim, reported by outlets that also did not read the board. Treat it as a vendor figure.
Where the training data came from — an actual answer
This is the part that makes Craft unusual, and it is worth being precise about what is established and what is not.
Canva did not build Craft on scraped web images. It built it on its own template library, which was made by its contributor community — and Canva has run a programme around exactly that question since 2023. At launch Canva announced a US$200 million fund over three years, with its CEO stating the company would use creators’ work for AI training only if they opt in. Opted-in creators receive an initial payment followed by monthly payments for continued use. A separate programme pays template royalties when a design that referenced a creator’s template is published — not merely when someone selects it.
Canva’s contributor agreement was updated effective 14 January 2026, clarifying that AI-generated content is not permitted in the programme and giving Creator Brands AI/ML opt-out controls in their Privacy Settings as well as through the older form.
Two honest caveats. First, Canva’s help pages and newsroom return 403 to automated readers, so we are relying on Canva’s own announcement as reported and on The Register’s contemporaneous 2023 coverage of the fund; we could not read the current terms directly. Second, an opt-in programme with payments is a structure, not a guarantee — we have not seen figures for what proportion of contributors opted in, or what a monthly payment actually amounts to.
But set it against this month’s other releases. Bria’s FIBO 2.0 claims 100% licensed training data and then gates the weights under a non-commercial licence. Getty Images filed for Chapter 11 while still selling “permissioned content” AI, with its Class A holders getting nothing. Canva is the one that can name the mechanism, point to a dated agreement, and say who gets paid when.
That does not make Craft better at making pictures. It makes it the one you can put in front of a client’s legal team without flinching.
What to actually do with it
If you already pay for Canva, Craft is sitting in Dream Lab now and the layout-first behaviour is immediately visible on anything with type in it — try a poster brief with a headline, a subhead and a price, which is where general-purpose models fall apart. If you are building something, the public API is the interesting surface, and the MCP integration is the one to watch if you want generation inside a chat workflow.
If you are evaluating it against Midjourney or a Flux variant for illustration work, Craft’s training corpus cuts both ways: a model taught on templates is good at templates, and templates are not photographs. That is a design decision, not a defect, but it sets the range.
Sources
- Canva releases its first image generation model — TechCrunch, 22 October 2026
- Canva Craft — Canva newsroom
- Canva launches Craft, its first in-house image model — The Decoder
- Canva creates $200m kitty for creators modelled by its AI — The Register, 2023
- Introducing Magic Layers — Canva newsroom
- DesignArena