Canva’s first in-house image model plans a design before drawing

A model launch makes Canva’s research organization visible

On October 22, Canva introduced Craft, its first image-generation model built and trained from scratch by Canva Research. The company says Craft is available in Dream Lab for Canva users and through a public developer API. TechCrunch independently reported the launch and described the same two routes. That makes the news more than a laboratory reveal: Canva is putting a model developed inside its design platform into a product that customers and developers can try.

Canva’s newsroom also introduced the research group behind the work. The company says Canva Research now has more than 160 AI scientists and engineers globally, within a broader AI unit of 400 people. Canva also says the group came together in 2025 after prior acquisitions, including Kaleido and Leonardo, and later brought in MangoAI and Simtheory. Those numbers and the group’s history are company disclosures, not an independently audited headcount.

Canva’s explanation for building its own models is tied to its existing product. It says it wants models that understand the structure of posters, presentations, social posts, and other designs, and that it can combine its own models with partner models. This is a different development strategy from a general image generator sold only through a separate API: Canva can train for its own editing tools and ship the result inside the workflow where users already arrange text and objects.

Craft separates composition planning from image rendering

Canva describes Craft as a two-stage model. A planning model interprets the prompt and produces a composition plan: what elements should appear, where they should sit, how large they should be, and what text belongs in the image. A separate rendering model turns that plan into an image. Canva says the output can reach 2K resolution in more than 30 aspect ratios, which is useful for switching between a poster, a slide, and a social graphic.

This design-first approach targets a common image-model problem. A picture can look polished and still fail as a usable design if the headline is unreadable, a product sits in the wrong place, or there is no room for a callout. Canva says its model treats visual hierarchy and text placement as part of the plan before drawing. That is a product claim, not proof that every prompt will produce legible copy or a correctly composed layout.

The technical interface adds a more concrete test. Canva’s developer announcement says Craft’s API model value is craft_v1_2k; it supports prompts in any language and image text in 55 languages. A developer can pin that model for a request, but the same note says a generation may fail if Craft is unavailable at that moment. That makes API access useful for testing, while leaving reliability and latency to be measured in a team’s actual workflow.

Canva says it trained on designs, not only photographs

Canva says in its Research announcement that Craft’s training emphasized millions of human-made templates from its content library, including posters, social posts, presentations, infographics, menus, and typography-heavy layouts. The company’s argument is that templates teach the model how hierarchy, spacing, and lettering work together, whereas a collection of flat photographs provides less direct experience with the structure of a finished design.

That explanation helps readers understand what the model was built to optimize, but it does not independently establish where Craft will outperform other image systems. The launch announcement does not provide a complete training dataset, per-category error rates, or a model card with technical details comparable to a research paper. It also does not explain contributor-level data choices in the launch post. Those are limits of the available announcement, not evidence by themselves of a problem with the training data.

The company says Magic Layers can turn a generated image into editable design elements. It also says future editing and reference-image capabilities are still being developed. So “editable” describes the broader Canva workflow and a design direction; users should check which output elements are actually separable in the version they use, rather than assume every object and word is independently editable from the first generation.

The launch reaches the editor and developer endpoint

Canva says in its developer forum notice that Craft is available globally in Dream Lab and through its Public API; the company’s developer forum names the image-generation endpoint and says each job uses the user’s AI credits. That means a team can evaluate the product in the place it already designs, or send prompts through an integration. The API is not an unlimited free quota, and the forum says a pinned Craft job can return a model-unavailable error if the service cannot run it at that moment.

Canva says Craft will also reach its MCP server, but its newsroom labels that integration as coming soon. TechCrunch similarly reported the plan to support third-party chatbots such as ChatGPT and Gemini. For a developer choosing a route today, the current options are Dream Lab and the public API; the MCP route is announced, not yet available. In these paths Craft is called as a hosted Canva model; the reviewed materials do not offer downloadable weights for local hosting.

A leaderboard result is a useful signal, not a full product evaluation

Canva says Craft debuted fourth overall on DesignArena and first in People & Portraits and Landscapes & Nature; it uses those results to call Canva the second most advanced image-generation lab. DesignArena describes itself as a crowdsourced benchmark in which real users compare generated outputs and choose a preference. That can show which image people favor in a particular matchup, but it is not a direct measure of how accurately a model follows a brand guide, how editable its files are, or what it costs a business to use.

Canva also says its designers compared more than 25,000 image pairs in internal evaluations. That is useful evidence of substantial internal testing, but the public post does not release the prompt set, participant agreement, or category-level result tables needed to reproduce the findings. TechCrunch confirms the release and availability, but did not independently test Craft. The sharpest current statement is therefore that Canva shipped a design-focused model and shared company-generated evaluation claims, while the product’s everyday reliability still needs outside and user testing.

The model’s value will depend on the whole design loop

Canva’s research lab is interesting because it connects image generation to the rest of a design process: planning, rendering, text placement, layered editing, and publication. A company that owns both the model and the editing environment can tune them together. That might reduce the number of manual corrections, but the launch material does not yet show how much time Craft saves against an existing generator plus a human designer.

For a team deciding whether to use Craft, give it and the current tool the same brief. Compare text accuracy, placement, brand consistency, editability, and rework; this shows whether the design-specific approach helps with a particular poster, campaign, or slide deck. For API teams, separately test image text across the languages they use and the client’s handling of model-unavailable errors. Canva documents language support and a possible unavailable-model failure, but offers no comparative result for those workflows.