Canva releases Craft, its first image generation model
Canva Craft plans a layout before rendering it, outputs up to 2K in 30 aspect ratios, and was trained on human-designed Canva templates rather than photographs. Results can be split into layered, editable designs, and the model is available in Canva's Dream Lab and API.
Editable layers are the real feature. A flat generated image is a dead end the moment a client wants the headline moved; a layered one can enter a normal design revision cycle. Expect video tools to be judged the same way — not on how good the output looks, but on how much of it you can still change.
Canva building its own image model is less interesting as a challenge to Google, OpenAI, ByteDance or Meta than as a statement about what a design platform thinks generated output should be. Craft was trained on designs rather than photographs — posters, social posts, presentations, infographics, menus and typography-heavy layouts drawn from human-made templates in Canva’s library — and it plans before it draws. That is a different idea of an image: a composition with parts, not a picture.
TechCrunch describes a two-stage process. The model first interprets the prompt and produces a plan covering element placement, relative scale, visual hierarchy and the text to be rendered, and only then generates the image. Output goes up to 2K in around 30 aspect ratios. Canva says text comes out legible and that targeted edits leave the rest of the image untouched; results can be converted into layered, editable designs with Magic Layers, its image-to-design model. Craft is available to all users in the Dream Lab tab and through the Canva API, with support inside ChatGPT and Gemini via Canva’s MCP planned but undated, and two cheaper variants, Craftfast and Craft-ultrafast, in development. Researchers from Leonardo.ai, which Canva acquired in 2024, contributed. The article notes there are no benchmarks or independent evaluations; the performance claims are Canva’s.
The planning stage deserves as much attention as the layers. In campaign work, the expensive moment for generated imagery is rarely the first frame; it is the tenth revision, when legal needs a disclaimer, the client wants the logo larger, a market needs a localised headline and the media plan adds three formats. A flat image means regenerating and hoping, or retouching by hand. A design that arrives with text as text, elements on separate layers and an explicit hierarchy can go straight into ordinary resizing and versioning. A plan is also an inspectable intermediate: an art director can, in principle, check the hierarchy before anything is rendered. It is the same logic as the Mise-en-Video paper this month — decide the structure, then make the image.
Provenance is where we would ask questions before using it on a high-profile campaign. “Trained on designs, not photographs” is a narrower statement than “trained on licensed material.” The templates were made by people, and the article does not say on what terms contributors made them available or whether those terms covered model training. A brand legal team will want that answered in writing. There is also a creative question about whether a model trained on templates pulls outputs toward a template look, which is exactly what brands with a distinct identity are paying to avoid.
The business context explains some of the timing. Canva has cut its 2026 growth forecast, citing high AI costs, and has relied on several third-party image models; owning one should lower its API spend. That is a reasonable motive, and it means Craft’s roadmap will be shaped by cost as well as quality. The signals to watch are pricing, independent comparisons, and whether layered output survives the MCP route into chat tools — a layered design flattened back into a single image inside a chatbot would lose the one property that makes it different.