On October 22, Canva released Craft, its first image generation model built in-house (TechCrunch). Canva's launch post on X passed 1.8 million views within two days, and most of the coverage focused on one design choice: Craft plans an image before it draws it.

This post explains what that means, why it matters for posters, menus and social graphics, and how you can use the same plan-first approach in your prompts today with the models on EasyImagen.

What Canva Craft actually is

According to TechCrunch and TestingCatalog, here are the main facts:

  • Two stages. A planning step first works out the elements, how they sit relative to each other, their scale, the visual hierarchy, and the text to render. A separate renderer then produces the image.
  • Trained on designs, not photos. Canva says it trained the model on human-made templates from its library: posters, social posts, presentations, infographics, menus and other layouts with a lot of type.
  • Output. Up to 2K resolution, more than 30 preset aspect ratios plus custom ones, and text rendered in 55 languages.
  • Availability. Craft is in Canva's Dream Lab for all users and in the Canva API. Support for AI assistants through Canva's MCP is planned, and two cheaper variants, Craft-fast and Craft-ultrafast, are "coming soon".
  • Rankings. TestingCatalog reports that Craft debuted at #4 overall on DesignArena, and at #1 in both People & Portraits and Landscapes & Nature.

Canva also says the plan includes bounding boxes for each element. That lets its Magic Layers tool turn a generated image into separate, editable text and objects. Most image models can't do this today.

Why "plan first" matters for design images

Most text-to-image failures in design work have nothing to do with how good the picture looks. They are layout failures:

  • The headline is misspelled or cut off at the edge.
  • The date ends up larger than the title.
  • An extra line of fake text appears in a corner.
  • Fixing one element means regenerating the whole image and losing the parts you liked.

A model that settles the layout before rendering pixels is better placed to avoid all four. You can't run Craft's planner inside another model, but you can do the planning yourself and write it into the prompt. Modern models like GPT Image 2 and Nano Banana follow a clearly structured brief very closely.

A plan-first prompt template

Write your prompt in roughly the order a designer would build the layout:

  1. Format. Portrait poster, menu card, square post, plus how much margin to leave.
  2. Zones. What goes at the top, middle and bottom, one line per zone.
  3. Exact text. Every string in quotes, exactly as it should appear.
  4. Hierarchy. Which element is largest, second and third.
  5. Style. Medium, palette and texture, kept short.
  6. A closing rule. "Render exactly the text above, spelled exactly, and no other words."

The closing rule matters more than it looks. Without it, models tend to fill empty space with small, fake text.

Example 1: an event poster with GPT Image 2

Design a print poster for a fictional music event. Plan the layout first, then render it.

Layout plan:
- Format: portrait poster with generous margins.
- Top third: the headline "RIVERSIDE JAZZ NIGHTS" in a tall condensed serif, cream on deep navy, set on two lines.
- Middle: one illustrated upright bass leaning against an old lamppost beside a river at dusk, flat screen-print style in mustard and teal.
- Lower band: three columns of small caps: "FRI 6 NOV" / "SAT 7 NOV" / "SUN 8 NOV", and below them one centered line: "Old Harbour Pier · Doors 7 PM".
- Bottom edge: a thin rule and the line "Free entry · All ages".

Visual hierarchy: headline largest, illustration second, dates third. Render exactly the text above, spelled exactly, and no other words.

Riverside Jazz Nights poster with an upright bass by a lamppost at dusk and three event dates

This is the first and only generation, at a 2:3 ratio. Every string is spelled correctly. The three dates sit in the three columns we asked for, the headline takes the top third, and no stray text appears. The model even added the dividers between the dates, which fits the "three columns" instruction.

Example 2: a menu card with Nano Banana

Menus are hard for image models because they combine many short strings with alignment. That is the kind of template Canva says Craft learned from.

Create a one-page cafe menu card. Plan the layout before drawing.

Layout plan:
- Header: "MORNING FIELD CAFE" centered at the top, with a small wheat sprig drawing above it.
- Two equal columns below the header. Left column titled "COFFEE": Espresso 3.00, Flat White 4.20, Oat Latte 4.80. Right column titled "BAKERY": Butter Croissant 3.50, Cardamom Bun 4.00, Banana Bread 3.80.
- Thin dotted leaders between each item and its price.
- Footer: "Open daily 7 AM to 3 PM".

Style: warm off-white paper with letterpress texture, terracotta and forest green ink, flat top-down view, the card fills the frame. Render only this text, spelled exactly.

Cafe menu card with Coffee and Bakery columns, dotted price leaders and a wheat sprig

All six items and prices came out correct and aligned, with the dotted leaders in terracotta. One instruction was only partly followed: the card doesn't fill the frame, and a white margin shows around it. If you need an edge-to-edge result, crop the image or say "no background visible" in the prompt.

Tips for getting text right

  • Keep the copy short. Two or three lines of small text work reliably. A full paragraph does not.
  • Quote every string, and write numbers and dates exactly as they should appear.
  • Name the hierarchy explicitly. "Headline largest, illustration second, dates third" stops the model from guessing.
  • Use one focal illustration. Busy scenes compete with the type.
  • If one word comes out wrong, shorten that line instead of adding more instructions.

Where Craft still has the edge

A prompt can describe a plan, but the result is still a flat image. Craft's advantage is that it hands its plan to Canva's editor. Through Magic Layers, the headline becomes real, editable text, and you can move an element without regenerating the whole image. If you need a layered file you can keep editing, that is a real difference. If you need a finished poster, menu or flyer in one pass, plan-first prompting closes most of the gap.

Try it

Copy either prompt above into the EasyImagen generator and swap in your own event, dishes or dates. For a ready-made seasonal version of the same idea, see the AI Halloween Poster Generator. For a matching social post, try this Halloween cat costume photo prompt or browse more prompt ideas.