Canva Craft plans a design before it draws it: what builders should note

What happened

Canva has launched Canva Craft, an image generation model it built itself. According to The Verge, it competes directly with image models from OpenAI, Microsoft and Google. The headline design choice is a two-stage architecture: the model plans a design first and generates the image second. Canva says this allows "greater precision, consistency and control."

This follows Canva's broader push into its own models. Coverage of its April 2026 Canva AI 2.0 announcement, such as Constellation Research's write-up, describes an in-house model program that includes named models like Lucid Origin and Proteus. Craft is the newest entry we have found.

Why planning first is the interesting part

Most text-to-image systems turn a prompt straight into pixels. That works for a single illustration. It struggles with design work, where a poster or slide has to respect layout, hierarchy, text placement and consistency across variants. Putting a planning step ahead of rendering gives the system an intermediate artifact that can be constrained, inspected and reused.

We have not seen a technical paper for Craft, so we can't say what the plan looks like internally. Treat the architecture description as Canva's own framing for now.

The pattern builders already know

If you build with LLMs, this should look familiar. Reliable agent systems usually split deciding from doing: a planner produces a structured intent, and an executor carries it out. The structured step is where you can validate, log and reject things before anything expensive or irreversible happens.

• Inspectable intermediate state: a plan can be checked by code or a human, while raw pixels or free-form text cannot.
• Consistency: reusing one plan across variants is easier than hoping repeated sampling agrees with itself.
• Targeted edits: changing the plan and re-rendering beats re-prompting from scratch.

What to watch

The open questions are practical ones. Is the plan exposed to users or APIs, or hidden inside the product? Can it be edited? How well do outputs hold up against brand constraints? Until those are answered in documentation or independent testing, the claim of more control is a hypothesis rather than a measured result.

For product teams the lesson doesn't depend on Craft's quality. If your generative feature fails on consistency, adding a structured planning stage is usually a better first move than a bigger model.