Canva Craft Is Here: Canva's First In-House Image Model Plans the Picture Before It Paints It

By Muzamil Ahmad · Updated 2026-10-27 · Image

#image#canva craft#canva ai image generator#canva craft dream lab#ai image generator design#canva craft vs midjourney#text to image ai

Save

How we write our guides: we start from the vendor’s own announcement or documentation, not from other “best of” lists. Launch dates and features are checked against that primary source; anything reported only by coverage is labelled as reported. Prices and allowances are given only when the vendor publishes them — otherwise we say “not published” rather than guessing. We do not invent test results, prices, or rankings.

# Canva Craft Is Here: Canva's First In-House Image Model Plans the Picture Before It Paints It

Let us start with the honest framing, because this article is not breaking news. Canva Craft launched on 22 October 2026 — five days before this guide — inside Canva's Dream Lab and through Canva's public API, according to Canva's own announcement. It is not a leak, not a rumour, and not a waitlist announcement. It is a shipped model from a company used by more than a quarter of a billion people, and it is worth a careful look precisely because it is already in your Canva account, whether or not you noticed it arrive.

The headline idea is simple, and it is the reason this model is interesting beyond the usual launch-week noise. Most image generators turn a prompt straight into pixels. Craft does not. It first writes a plan for the picture — what goes where, how big each element is, what should catch the eye first, and what text has to appear — and only then paints the picture from that plan. Canva calls this a two-stage architecture, and the whole pitch for Craft rests on it: if the model reasons about composition before it renders, the argument goes, it should place objects where you asked, spell the words on the poster correctly, and let you change one element without redrawing everything else.

That is a claim, not a proven fact about your next poster. Canva's own announcement supports it with a leaderboard debut and with category results, and independent coverage has already shown how fast those leaderboard numbers move. This guide separates what Canva has confirmed from what is merely reported, what is live from what is still promised, and what Craft is actually good for from what it merely sounds good for. If you design social posts, posters, menus, slides, or marketing assets — the exact work Canva was built for — Craft is aimed squarely at you. If you want photorealistic photography, fine-art illustration, or a model you can run on your own machine, the answer is more complicated, and this guide says so plainly.

What Canva Craft Actually Is

Canva Craft is an image generation model built specifically for design work. That sentence carries more weight than it first appears, because "built for design" is doing two jobs at once: it describes what the model is for, and it explains why Canva built it at all.

Until now, Canva leaned on other companies' models for image generation inside its editor. Dream Lab, Canva's AI image tool, launched in October 2024 running on Phoenix, the model from Leonardo.ai, the Sydney start-up Canva acquired that year. Launch coverage reports that researchers from Leonardo worked on Craft, with Canva telling TechCrunch so — that detail comes from coverage, not from Canva's own announcement, and this guide marks it as reported rather than confirmed. What Canva itself confirms is the direction of travel: Craft is the first foundation model built and trained from scratch by Canva Research, and it moves Canva from renting other companies' image models to running its own. That shift matters for cost, for quality control, for speed, and for how fast Canva can change the model when designers ask for something different. It is also the reason Canva could build the model around its own template library rather than around somebody else's dataset.

The scale of the audience is part of the story too. Canva says more than a quarter of a billion people use its platform to create over 400 designs every second. A model that serves that many people cannot just be good; it has to be fast and cheap enough to run at that volume. Canva says it used advanced model distillation to cut the number of inference steps Craft needs to a fraction of the original model, and built a custom inference engine optimised for the planning stage, making Craft an order of magnitude faster than its first internal version. Those engineering claims are Canva's own — no independent test has verified the speed figures — but they explain why the faster variants are the next step rather than an afterthought. A model built for 400 designs a second has to be engineered for throughput from day one.

One more piece of reported context, clearly labelled as such: coverage of the launch describes Craft as built from scratch in about six months by a roughly 160-person Canva Research team, quoting co-founder Cliff Obrecht. Canva's own announcement does not state a team size or a build time, so treat the six-month, 160-person figures as reported, not confirmed. They are included here because they help explain how a design company shipped a foundation model this quickly — a large team with a ready-made dataset advantage — but they are not facts Canva itself published in its announcement, and this guide does not present them as such.

The Two-Stage Architecture: Plan First, Paint Second

Here is the idea at the centre of Craft, in Canva's own description. Conventional image generation models generate pixels directly from a prompt. Craft separates compositional planning from image rendering, so the model reasons about the structure of an image before generating it.

In the first stage, an image planning model interprets the prompt and writes a detailed plan of the composition: which elements appear, their spatial relationships, their relative scale, the visual hierarchy, and any text that needs to appear in the image. In the second stage, a dedicated image rendering model translates that plan into a high-fidelity image at up to 2K resolution across more than 30 aspect ratios.

Why should you care? Because the failures of ordinary image generators are usually planning failures, not painting failures. The picture looks beautiful and the headline says the wrong thing. The product shot is gorgeous and the bottle is on the wrong side of the frame. The poster is striking and the date is misspelled, or the text floats somewhere it was never asked to be. A model that writes down where everything goes before it draws has, in principle, a chance to get those things right — and, just as importantly, a plan gives the user something to correct. Canva says this architecture enables more granular control over individual elements, allowing targeted changes without unnecessarily altering everything else. If that holds up in practice, it is the difference between regenerating a poster ten times hoping for a lucky roll and editing the one element that is wrong.

Canva illustrates the point with a deliberately simple example: ask for a purple sky, long shadows, and a lone tree, and you should get all three, with the emphasis where you wanted it. That is a modest claim, and modesty is a good sign. The strongest image models fail most often on compound requests — three or four requirements at once — because each requirement competes with the others in a single pass from prompt to pixels. A planning stage is a way of holding all the requirements in view at once. It does not guarantee success, and this guide has not independently tested Craft against a battery of compound prompts, so it makes no claim about a success rate. But the mechanism is sound, and it is the right thing to test first when you try Craft yourself: give it a prompt with a headline, a subheading, a product position, and a background mood, and check whether all four survive into the finished image.

There is a second, less obvious benefit to the two-stage design. A plan is structured information, and structured information converts more cleanly into editable layers than a flat grid of pixels does. Canva says Craft uses JSON boundary boxing to make it easier to convert images into editable layers with Magic Layers, with more accurate and faster results. That is the bridge between "generate a picture" and "finish a design," and it is where Craft's design-first training and its planning architecture meet. A picture you cannot edit is a dead end for a designer; a picture that arrives as layers you can move, retype, and recolour is a starting point.

Trained on Templates, Not Photographs

The training data is the other half of Craft's pitch, and it is the half that most clearly separates it from its rivals. While most image models are trained primarily on photographs, Canva says Craft's training places a much greater emphasis on design. Its training data includes millions of human-made templates from Canva's content library, spanning professionally designed posters, social posts, presentations, infographics, menus, and typography-rich layouts.

The logic is straightforward. A photograph teaches a model what the world looks like. A template teaches a model what a usable design looks like: how a headline sits against body copy, how hierarchy guides the eye from the most important element to the least, how spacing and alignment make a layout feel intentional rather than accidental. Unlike flat images, Canva says, these templates teach the model exactly how layout, hierarchy, and typography fit together. That is a genuine advantage for the work Canva's users actually do — and it is also a genuine limitation for work they do not. A model trained mainly on templates should be expected to be stronger on posters and menus than on photorealistic wildlife photography, and this guide would not recommend judging Craft primarily on the latter.

After pre-training, Canva says it optimised Craft through supervised fine-tuning and reinforcement learning, incorporating feedback from its professional designers to improve instruction following, typography accuracy, and overall visual quality. For internal evaluation, Canva's professional designers worked alongside dozens of researchers to assess more than 25,000 pairs of images in head-to-head comparisons with leading models, and where automated evaluations and human judgments differed, Canva says it sided with the designer. That last detail is worth noting. Automated image benchmarks reward things machines can measure easily; designers reward things humans actually notice. A lab that explicitly prefers designer judgment over automated scores is making a statement about whose opinion counts, and for a design tool, that is the right statement — even though it also means Canva's internal evaluations cannot be independently reproduced from the announcement alone.

A question readers should keep in mind, because Canva's announcement does not answer it: the announcement does not publish a detailed dataset provenance breakdown, a model card, or a system card for Craft. It says the training emphasises human-made templates from Canva's library, but it does not say what share of the data those templates represent, what licensing or consent arrangements cover them, or what other data sources were used. Those are reasonable questions for any foundation model in 2026, and their absence is a gap, not a scandal. The honest position is that Craft's training story is plausible and well-matched to its purpose, but only partly documented in public so far.

What Craft Is Good For — And What It Is Not

Canva itself lists the work Craft is built to support, and the list is a useful map of its strengths: social content and marketing assets (posters, social posts, flyers, presentation visuals, ads) with layouts, hierarchy, and text ready to use; lifestyle and editorial photography; illustration and digital art in flat, vector, painterly, and hand-drawn styles; people and characters, including complex scenes that stay coherent; nature and environment imagery; and creative, conceptual work that brings real and stylised elements together.

Read that list with a designer's eye and a pattern emerges. The first category — social and marketing assets — is where the template training and the planning architecture reinforce each other most directly. A social post is a layout problem with a picture inside it, and Craft is built to solve layout problems. The people, portrait, and landscape categories are where Canva claims its strongest leaderboard results (see the leaderboard section below), and they are also the categories where a planning stage helps most: a group portrait that stays coherent, a landscape with a deliberate foreground and background, a product shot with the product where it belongs.

What Craft is not, at least not yet, is equally clear. It is not an open-weights model you can download and run yourself. Our FLUX 3 Image Klein guide covers a model built for exactly that use case — 4K images on your own GPU — and Craft is a different kind of product: a hosted model inside Canva's editor and API, not a checkpoint you own. It is not a video generator, not a 3D model, and not a replacement for a photographer or an illustrator on work where originality, licensing, or a human eye matter most. And it is not, on the evidence published so far, a proven winner on fine print and technical drawings: Canva itself says future work will improve "fine print and technical drawings," which is a polite way of saying those are current weaknesses. If your work lives in small text, dense diagrams, or precise technical illustration, test Craft carefully before relying on it.

It is also worth placing Craft against the broader image-generator landscape rather than treating it as a world unto itself. If you are choosing between free generators, our best free AI image generators guide compares the field — and until this guide, Canva appeared there only through Magic Media, its earlier in-editor image tool, not through a dedicated Craft guide. That gap is exactly what this article fills. If your comparison is head-to-head between the big names, our Midjourney vs DALL-E vs Stable Diffusion comparison covers the established rivalry that Craft now joins from a different angle: not as a better art machine, but as a design machine. And if your end goal is a finished brand asset rather than a picture, our best free AI logo makers guide is the right next read, because a logo still needs the kind of deliberate, editable construction that a raw image generator — Craft included — does not finish on its own.

How to Try Craft in Dream Lab

Trying Craft takes no setup beyond a Canva account. Canva says Craft is available starting 22 October 2026 to all Canva users globally. To start, open Dream Lab in the bottom-left panel of Canva, or follow Canva's own link from its announcement to Dream Lab. There is no separate Craft app, no waitlist for the base model, and no separate download.

Once you are in Dream Lab, the workflow follows the same shape as any image generator, with two Craft-specific choices to make. First, pick your aspect ratio: Craft offers more than 30, or you can enter a custom size, so choose the format for the place the image will actually live — a portrait story, a landscape banner, a square post, a poster — rather than generating a default square and cropping it later. Cropping after the fact throws away the composition the planning stage just worked out, which defeats the point of the model. Second, write the prompt the way the planner wants to read it: name the elements, their positions, their relative sizes, and the exact text, rather than describing a mood and hoping the layout sorts itself out. The planner can only plan what you tell it.

A practical prompt pattern that plays to Craft's strengths looks like this: state the format and purpose first, then the headline text in quotation marks, then the elements in order of importance, then the style. For example: "Portrait poster for a weekend farmers market. Headline: 'Saturday Market'. Subheading: 'Fresh, local, every week'. A wooden stall with vegetables in the foreground, a crowd softly out of focus behind, warm morning light." Every clause in that prompt is something the planning stage can act on: a format, two text strings with a hierarchy between them, a foreground element, a background treatment, and a lighting mood. Compare that with "a nice farmers market poster," which gives the planner almost nothing to plan with and leaves the model to guess. The difference between those two prompts is, in miniature, the difference Craft is trying to sell.

One honest limitation to carry into your first session: Canva's Dream Lab help material says some Dream Lab features use your AI allowance, and Canva has not said how many uses a Craft generation consumes from that allowance. The practical consequence is simple — check your own allowance before a big batch, and do not assume Craft generations are free and unlimited just because Craft itself carries no separate price tag. The model is available to all users, but "available" and "unlimited" are different words, and Canva has published the first, not the second.

The API: Craft for Developers

For developers and product teams, Craft is also available through Canva's public API, and this is where the launch gets interesting beyond Canva's own editor. Canva says developers can build Craft into their own tools — marketing platforms, e-commerce sites, ad builders, print services — and that the API is the quickest way to test Craft on your own prompts and compare it with other models.

Canva's developer community post for the launch gives the concrete details, via the Connect API announcement. Craft is a new model value, craft_v1_2k, on the existing POST /v1/image-generations endpoint: pin that value to generate with Craft, pass your prompt and aspect ratio, then poll the job until it succeeds. The download URL is valid for 24 hours unless you set the upload option to save the image to the user's uploads. The call requires a user access token with the asset:write scope. Craft supports the aspect ratios the endpoint offers, and if Craft is unavailable, the job fails with a model_not_available error — in which case you retry, or drop the model parameter and let Canva pick another model.

That API surface matters for a reason that goes beyond convenience. An image model that only lives inside one editor is a feature. An image model with a public API is infrastructure: it can sit inside a marketing platform that generates a campaign's assets, an e-commerce site that renders product lifestyle shots, or an ad builder that produces variants for testing. Canva is explicitly pitching that second role, and the API's existence on launch day — not months later — is a sign of how seriously it means it.

Two cautions belong here, both from Canva's own material rather than from scepticism. First, the developer post describes the API as a preview-style surface in places, and community discussion notes that breaking changes remain possible while the surface is young; build accordingly, with the model parameter pinned and a fallback in place, rather than hard-coding Craft as your only path. Second, and more importantly, there is no published API price for Craft. Jobs consume the end user's credits under their Canva plan, which means a developer building on Craft cannot yet forecast a per-image cost from a rate card. That is a real planning gap for anyone building a product on the API, and it is the single biggest reason this guide does not recommend building a business on Craft alone until Canva publishes pricing. Test it, prototype with it, compare it — but do not price a product around a number Canva has not given you.

Magic Layers: Where Craft Stops Being a Picture

If there is one feature that decides whether Craft matters to working designers, it is Magic Layers. Canva's announcement says the same underlying structure that makes Craft plan before it paints also makes its outputs particularly well suited to Magic Layers, which can transform generated images into fully editable design elements. The planning stage produces structured information about elements and boundaries; Magic Layers turns that structure into layers you can select, move, retype, and restyle.

Think about what that changes in practice. A traditional image generator hands you a finished picture. If the headline is wrong, you regenerate and hope. If the background colour clashes with your brand, you open an editor and paint over it. If the layout needs to shift for a different format, you start again. A generated image that arrives as editable elements behaves like a design file instead: the headline is text you can retype, the background is a shape you can recolour, the product is an element you can move. That is the difference between an image generator and a design tool, and it is the reason Canva trained Craft on templates rather than photographs in the first place. Templates are already layered, structured designs; a model that learned from them has a head start on producing outputs that decompose back into layers.

The honest caveat is that Magic Layers conversion is a Canva-side process, not a property of the image file itself. A Craft image downloaded as a flat file and opened elsewhere does not carry its layers with it. The editable benefit lives inside Canva's editor, which is exactly where Canva wants you — and that is fine, as long as you understand the trade. If your workflow ends in Canva, Magic Layers is a genuine advantage. If your workflow ends in another tool, Craft's output is, for you, a flat image like any other generator's, and you should judge it on image quality alone.

For readers whose work ends in other tools, two of our existing guides cover the handoff. Our how to remove a background guide covers the most common post-generation fix — isolating a subject from a generated scene — and our best AI presentation makers guide covers what happens when a generated visual still needs to become a finished deck, where Canva's own Magic Design already appears as the template powerhouse. Craft plus Magic Layers is Canva's attempt to collapse those steps into one place; whether it succeeds for your work is something only a real project will tell you.

The DesignArena Leaderboard: What the Rank Does and Does Not Mean

No part of this launch has generated more confusion than the leaderboard, so this section takes its time and keeps confirmed and reported claims visibly apart.

DesignArena is a crowdsourced leaderboard: human voters compare images from different models head-to-head, and the rankings reflect those votes rather than an automated benchmark. Canva's own announcement makes two distinct claims about it. First, Craft debuted as the fourth most advanced image generation model on DesignArena. Second, that debut makes Canva Research the world's second most advanced lab for image generation — because, in Canva's telling, the three models ahead of Craft belong to a single rival lab, so the lab ranking differs from the model ranking. Canva also claims category results from the same debut: number one globally for People & Portraits (out of 87 models) and for Landscapes & Nature (out of 85 models), number two for Abstract & Patterns, and number four for Products.

Those are Canva's claims about its debut, and this guide reports them as Canva's claims. The confusion in launch coverage came from collapsing the two rankings into one: several reports described Craft as "second-best," which is the lab claim wearing the model's clothes. Once you separate them — model fourth, lab second — the coverage largely agrees, and the apparent contradiction dissolves. One widely cited report put Craft fourth with a score of 1,329 on 23 October, behind three OpenAI models and ahead of Google's Nano Banana 2.1; another analysis, checking the leaderboard on the evening of 23 October, found Craft's category position had already shifted in at least one tab. Both observations point the same way: DesignArena is a living vote count, not a fixed exam result, and a debut position is a snapshot of one day.

Here is the decision this guide has made, and why. The task for this article required checking the live DesignArena leaderboard today and printing no rank unless the live primary leaderboard confirmed it. That check was attempted: the live leaderboard page at designarena.ai returns a JavaScript-rendered page whose text fetch yields only the page shell, not the ranking table, so today's live position could not be independently confirmed from the primary source by the means available. The leaderboard therefore gets no current rank claim in this guide — only Canva's debut claims, labelled as Canva's, and the dated coverage snapshot above, labelled as coverage. If you see a current rank quoted anywhere, check the date on it before trusting it: on a crowdsourced board, a rank without a date is barely a fact at all.

What the leaderboard does establish, even with that caution, is that Craft arrived competitive rather than embarrassing — fourth on debut, first in two categories, in a field of 85-plus models per category. What it does not establish is that Craft is the best model for your poster, your product shot, or your menu. Leaderboard voters judge images in isolation; you judge them inside a design, under a deadline, with a brand to protect. The right response to a fourth-place debut is to try the model on your own work, not to buy or dismiss it on a number.

Pricing: The Missing Number

This is the shortest section in the guide, and the most important one for anyone planning to rely on Craft. Canva has not published a price for Craft.

To be precise about what is missing: there is no published per-image API price, no published statement of how many AI-allowance uses a Craft generation consumes on a Free, Pro, or Business plan, and no price or date for Craft-fast or Craft-ultrafast. Canva's announcement says the faster variants will deliver high-quality generation "at a fraction of the cost of leading frontier models," but a fraction of an unnamed number is not a price. Third-party analyses have attempted to derive per-image figures from Canva's general plan allowances — dividing a plan's monthly AI uses by its price, for example — but those figures depend on an assumption Canva has not confirmed, namely which allowance tier Craft draws from, and this guide does not repeat them as facts. A derived number built on an unconfirmed assumption is a guess with arithmetic, not a price.

What you can plan around is this: Craft is available inside Canva's existing plans through Dream Lab, and through the public API against the end user's credits. If you are an individual designer, try it inside the plan you already pay for and watch your allowance. If you are a developer, prototype against it and wait for a rate card before committing a product's pricing to it. And if you read a review that quotes a Craft price, ask for the source: as of this guide's writing, Canva itself has not given one.

What Is Still Promised, Not Shipped

A launch-day guide that only lists what is live would mislead you, because Canva's announcement is explicit about a roadmap, and the roadmap items are the ones most likely to be mistaken for live features in casual coverage. Here they are, each marked with its true status.

Craft-fast and Craft-ultrafast: promised, not out. Canva describes them as two upcoming variants designed to deliver high-quality image generation at a fraction of the cost of leading frontier models, built on the distillation and custom inference work described above. No release dates, no prices, and no independent tests exist yet. When they land, they are the versions that will decide whether Craft is cheap enough to use at volume — but today, they are a promise.

Canva MCP for Craft, including ChatGPT and Gemini access: coming soon, not live. Canva's announcement says Craft "also connects to AI assistants and automated workflows through Canva MCP," and then labels Canva MCP itself as "Coming Soon." The vision is clear: an AI assistant could make a launch graphic, a product shot, or a social post as one step in a bigger task. Coverage describes planned integrations for assistants such as ChatGPT and Gemini through that MCP connector. None of that is live for Craft today. If you read that you can already generate Craft images from inside ChatGPT, you are reading the roadmap as if it were the release.

More advanced editing and reference-image capabilities: future releases. Canva says the planning architecture provides the technical foundation for more advanced editing and reference-image capabilities coming in future releases. No dates are given. Reference-image editing — generating from or editing with an existing image as a guide — is one of the most requested features in image generation, and its absence from the launch is worth noting if that is your workflow.

Fine print and technical drawings: acknowledged as future improvement. Canva says future work will keep raising the bar on quality, detail, and control "to improve things like fine print and technical drawings." Read that as a candid admission about current limits, and test accordingly.

Safety, Moderation, and the Questions That Remain

Canva says safety is built into every stage of how it develops and deploys AI, with continuous testing and rigorous safety evaluations covering brand safety, cultural context, and the potential for harmful content, and that Craft, like all Canva AI products, is backed by Canva Shield: input and output moderation, privacy and safety controls, and reporting tools for flagging misleading content. Those are Canva's descriptions of its own systems, and this guide reports them as such rather than as independently audited facts.

Beyond the safety systems, three open questions deserve a place in any honest guide. First, commercial terms: Canva's announcement does not publish detailed output licensing terms for Craft images in the way a model card might, and commercial users should check Canva's current terms for AI-generated content before using Craft output in client work. Second, data provenance: as noted above, the announcement describes the training emphasis on human-made templates but does not publish a full dataset breakdown. Third, evaluation transparency: the 25,000-pair designer evaluation is Canva's internal result, not a public benchmark you can rerun. None of these gaps is unusual for a launch-week announcement, and none of them is a reason to avoid Craft. All three are reasons to keep the distinction this guide has maintained throughout — confirmed versus reported versus promised — in mind as the model matures.

A Labelled Worked Example: One Campaign, Three Formats

To make the planning idea concrete without inventing a client or a case study, here is a clearly labelled fictional worked example. Imagine a fictional neighbourhood bakery — call it Maple & Crumb, invented for this guide, not a real business — launching a weekend market stall. It needs three assets: a square social post, a portrait story, and a landscape banner, all carrying the same headline, "Saturday Market," and the same subheading, "Fresh, local, every week."

In a traditional single-pass generator, that is three separate generations, and the headline, the stall, and the lighting may drift between them: three rolls of the dice, three chances for the text to misspell or the composition to change. The arithmetic of the risk is simple — if each generation has even a modest chance of getting one element wrong, three independent generations multiply the chances that at least one asset needs redoing.

Craft's pitch is that the planning stage changes that arithmetic. One plan — headline, subheading, stall in the foreground, crowd softly behind, warm morning light — rendered in three aspect ratios should hold the composition steadier across formats, because the same compositional decisions feed all three renders. That is the claim to test, and the test is cheap: run the same brief in Craft's three formats, lay the results side by side, and check whether the headline spelling, the stall position, and the light survive all three. If they do, Craft has saved you the most expensive part of campaign production, which is not generation but consistency. If they do not, you have learned something valuable about the model for the price of three generations. Either way, you have an answer grounded in your own work rather than in a leaderboard.

The same logic scales to the other end of the size range. A small business producing a menu, a flyer, and a social post for the same promotion faces the same consistency problem, and a model trained on menus and flyers has a head start that a photograph-trained model does not. That is Craft's home ground, and it is where this guide expects it to age best.

Who Should Try Craft First — And Who Should Wait

Try Craft first if you design in Canva already, if your work is posters, social posts, menus, slides, flyers, or ads, if legible text on the image matters as much as the picture, or if you want generated images you can edit as layers inside Canva rather than flat files. Try it through the API first if you build marketing, e-commerce, ad, or print tools and want to compare a design-trained model against the general-purpose models you already use — but prototype, do not price, until Canva publishes a rate card.

Wait, or test more sceptically, if you need reference-image editing, if your work depends on fine print or technical drawings, if you need to run a model on your own hardware, or if you need Craft inside ChatGPT or Gemini today — that integration is coming soon, not live. And if your work is photorealistic photography or fine art, compare Craft against the specialist models in our best free AI image generators guide before assuming a design-trained model is the right tool for a photography job.

For social teams specifically, one more link is worth following. A generated image is only the middle of a social workflow — it still needs scheduling, captioning, and distribution — and our best AI social media tools guide covers that surrounding workflow. Craft improves the image step; it does not replace the rest of the job. Similarly, if your generated portraits are destined for profile pictures rather than posters, our best free AI headshot generators guide compares tools built for that narrower, more demanding task, where a face has to look like a specific person rather than a plausible stranger.

The Verdict

Canva Craft is not the most dramatic image-model launch of 2026, and that is a compliment. It is a disciplined one: a design company building its own model, trained on its own design library, architected around the actual failure modes of design work — misspelled headlines, misplaced objects, uneditable output — and shipped inside the editor where its users already work, with an API on day one. The planning architecture is a real idea with a real mechanism behind it, not a marketing metaphor. The template training is a genuine dataset advantage for posters, menus, and social assets, even if it is only partly documented in public. And Magic Layers is the feature that turns a picture into a design, which is the only transformation that matters to Canva's core user.

The gaps are equally real, and they are the gaps this guide has refused to paper over. No published price. No live leaderboard rank confirmed today. Faster variants and assistant integrations promised, not shipped. Reference editing and fine-print quality still on the roadmap. A training story that is plausible but not fully documented. None of those gaps makes Craft a bad model; all of them make it a young one, five days old at this guide's writing, and young models should be tried on real work, not believed on launch copy.

So try it the way this guide suggests: one real brief, three formats, side by side, checking whether the headline, the hierarchy, and the composition survive. If they do, you have found the fastest path from an idea to a finished, editable design that Canva has ever offered. If they do not, you have spent a few generations learning exactly where the planning promise meets its current limits — and that knowledge is worth more than any leaderboard number, including the ones in this article.