How to Generate AI Images with Canva Craft (2026)

On October 22, 2026, Canva launched Craft, the first image-generation model the company has built and trained entirely in-house. It is available to every Canva user right now through Dream Lab, and it changes the daily design workflow in a way most model launches do not. Instead of producing a flat picture you drop into a design, Craft generates images that are meant to be edited afterward: readable text, deliberate layouts, and structure that Canva's Magic Layers can break apart into elements you can actually touch.

This tutorial walks through the whole workflow step by step: where to find Craft inside Canva, how to write prompts that play to its layout-first strengths, how to turn a generation into a genuinely editable design, and where the honest limits are. No design degree required — if you can describe what you want in a sentence, you can use it.

What is Canva Craft

Until now, Canva's AI image tools ran on models from other companies — when Dream Lab launched in October 2024, it ran on Phoenix, the model from Leonardo.ai, the Sydney startup Canva acquired that same year. Craft is the first model built and trained from scratch by Canva Research, the company's in-house AI lab. According to Canva, researchers from the Leonardo.ai acquisition contributed to the work, and the team now numbers more than 160 researchers and engineers.

Craft's headline trick is a two-stage architecture. A planning model first reads your prompt and writes a detailed composition plan — elements, spatial relationships, sizes, hierarchy, and any text that must appear. A separate rendering model then turns that plan into a picture at up to 2K resolution, in more than 30 aspect ratios or a custom size. Canva says planning first keeps text legible, places objects where you asked, and lets you edit one element without rebuilding the whole image.

Two more facts shape how you should use it. First, Craft was trained with design, not photography, as the emphasis: millions of human-made templates from Canva's library — posters, social posts, presentations, infographics, menus, typography-rich layouts — followed by supervised fine-tuning and reinforcement learning guided by its professional designers, who compared more than 25,000 image pairs against rival models. Second, the benchmark story is entirely company-reported: Canva says Craft debuted fourth overall on DesignArena, with first-place finishes among 87 models for People & Portraits and among 85 for Landscapes & Nature. Those are Canva's numbers, not an independent verdict — though the template-trained design angle is genuinely different from what rivals ship.

What you need before you start

Almost nothing. Craft is available to all Canva users globally — per Canva's announcement, there is no waitlist and no separate product to install. A free Canva account gets you in. Inside Canva, Craft lives in Dream Lab, which you open from the bottom-left panel of the editor. Developers can reach the same model through Canva's public API: a staff post on Canva's developer forum explains that the Connect API image-generation endpoint accepts a pinned model value, craft_v1_2k, and that each request takes noticeably longer than other models — roughly 50 to 60 seconds end to end. Canva also says MCP support, which would let AI assistants create design-ready images inside larger workflows, is coming soon.

For this tutorial, everything happens in the Canva editor, which means the steps below work the same whether you are on the free or a paid plan.

Step-by-step: generate your first Craft image

Step 1: Open Dream Lab. In the Canva editor, look at the bottom-left panel and open Dream Lab. This is Canva's dedicated AI image space — the same surface that previously ran third-party models — and since October 22 it runs Craft.

Canva screenshot

Step 2: Write a layout-first prompt. This is the single most important step with Craft. Because the model plans composition before rendering, prompts that describe the layout get the best results. Name the format, the headline, and where things go: "A concert poster, headline at the top in large type, three act names listed below in two columns, a spotlight-lit stage photo behind everything, footer strip with the venue address." If you want specific words to appear in the image, put them in quotes exactly as you want them printed — the planner treats quoted text as text that must appear, and Craft was explicitly trained for typography accuracy. Prompts can be written in any language, and on-image text is supported in 55 languages, so write in whatever language you think in and quote the printed text in the language it should appear.

Step 3: Pick an aspect ratio. Craft supports more than 30 preset aspect ratios plus a custom size. Choose the one that matches the destination — a square for feed posts, portrait for stories or posters, landscape for banners or slide headers. Getting this right before generating matters: Craft's layout planning is tied to the canvas shape, and generating at the final ratio keeps text and composition where you meant them instead of cropping later.

Step 4: Generate. Run the generation. The planning stage maps out your elements first, then the renderer produces the image at up to 2K resolution. Check the result against your prompt: is the headline where you asked, is the text spelled correctly, do the objects sit in the right relative positions? With layout-first prompts, the first generation is often close; when it misses, the miss is usually a prompt detail, not a model lottery — tighten the wording and regenerate rather than rewriting everything.

Step 5: Refine with targeted edits. Craft's architecture is designed for granular edits: you can change one element — a purple sky, long shadows, a lone tree, as Canva's own examples put it — without unnecessarily altering everything else. Describe the change against the current image: "keep everything the same, but change the background to dusk" beats "make a new poster at dusk." This is where the two-stage design pays off in practice; edits stay consistent because the underlying plan is reused.

Step 6: Break it into layers with Magic Layers. Here is the step no other image generator gives you in the same tool. Canva's Magic Layers can convert a Craft-generated image into fully editable design elements — text becomes selectable text boxes, objects separate from the background. Craft's outputs are especially suited to this because the planner's internal structure, including JSON boundary boxing, maps cleanly onto editable layers, and Canva says the conversion is faster and more accurate on Craft images. After conversion, you can reword a heading, swap a font, recolor an element, or move objects around as if you had built the design by hand.

Step 7: Export or keep designing. Use the generated image as-is, or continue working in the Canva editor with the layered version — the point of the whole workflow is that generation and design are no longer two separate tools. Social graphics, posters, presentation visuals, ads: anything that used to start from a stock photo can now start from a generated image that arrives already editable.

Prompt patterns that play to Craft's strengths

Four patterns consistently get more out of the model:

Write a design brief, not a mood board. "Warm minimal coffee shop logo poster, name in bold sans-serif centered top, tagline underneath, line-art cup icon left of the name, muted beige background" outperforms "cozy coffee vibes." The planner rewards explicit structure.
Quote printed text exactly. Whatever is inside the quotes is what the typography stage tries to render. Keep on-image text short and put long copy in the design editor afterward instead.
Lean into its training. Craft saw millions of human-made templates, so it understands posters, infographics, menus, slide layouts, and social cards better than it understands pure photography. If your task looks like a template, say so in the prompt.
Keep long copy for the editor. Craft handles short quoted headlines well; put paragraphs and fine print into Canva text boxes after generation instead of inside the prompt.

Limits and honest caveats

A tutorial should also say what the tool cannot do yet. The DesignArena rankings are Canva's own announcement figures — first place in two categories sounds impressive, but leaderboard scores shift as new models arrive, and independent evaluation has not caught up with the launch claims. Advanced editing tools and reference-image capabilities are announced for future releases; today you can request targeted edits, but not yet work from a reference image the way some rival tools allow. The cheaper Craft-fast and Craft-ultrafast variants have been announced with no release dates or prices, and per the developer forum, API generation currently takes roughly a minute per image — noticeably slower than the models Dream Lab used before. Partner models will continue to run alongside Craft inside Canva, so do not assume every AI feature in the editor is Craft-powered. Finally, like all of Canva's AI products, Craft runs under Canva Shield: prompts and outputs are moderated, with privacy controls and a reporting channel for harmful or misleading content — which means certain prompts will be refused, by design.

Bottom line

Craft is the rare model launch that matters to non-technical users, because it lives where people already design. The practical move is to test it on one real task you do often — a poster with text, a social graphic, a menu — and judge it on layout fidelity, not photorealism. That is the dimension it was trained for, and the one its rivals mostly treat as an afterthought. If the headline lands in the right place and the text reads correctly on the first try, you have felt exactly what the two-stage architecture is for: an image that arrives as the beginning of a design, not the end of one.

Sources

Canva official newsroom, "Introducing Canva Craft" — canva․com
TestingCatalog, "Canva launches Craft image model with 2K output and API" — testingcatalog․com
BetaNews, "Canva launches Craft, its first in-house AI image model" — betanews․com
Canva Developers Community, "Connect API: Canva Craft is now available for image generation" — community․canva․dev
madrobot.blog, "Canva Craft: Canva launches its own AI image model in Dream Lab" — madrobot․blog