Canva releases Craft, its first in-house image generation model

What Canva released on 22 October

On 22 October 2026 Canva introduced Canva Craft, the first image generation model built and trained by Canva Research, the company's in-house AI lab. Canva calls it a foundation model trained from scratch rather than a fine-tune of someone else's system, and it is available immediately to every Canva user through the Dream Lab panel, as well as to developers through Canva's public API.

The model generates images at up to 2K resolution in more than 30 preset aspect ratios, and users can also type in a custom ratio. Canva says prompts can be written in any language and that Craft can render text in 55 languages. Two cheaper variants, Craft-fast and Craft-ultrafast, are announced but not yet released.

TechCrunch reported that Canva built the model to compete with image systems from Google, OpenAI, ByteDance and Meta, and that researchers from Leonardo.ai, the media-generation company Canva bought in 2024, contributed to it. TechCrunch also noted that Canva has used a number of different image models for generation in its product.

How the two-stage design works

The most concrete technical claim in Canva's post is about structure. Conventional text-to-image models go from a prompt straight to pixels. Craft splits the job in two. A planning model first reads the prompt and writes a detailed plan of the composition: which elements appear, where they sit relative to each other, their relative scale, the visual hierarchy, and any text that needs to appear in the image.

A separate rendering model then turns that plan into the final image. Canva argues that deciding the layout and the typography before drawing anything is what lets Craft produce legible text, put objects where they were asked for, and keep the relationships between elements consistent. It also says the split allows targeted edits, so a user can change one element without the model redrawing everything else.

Speed was a stated engineering problem. Canva says more than a quarter of a billion people use the product and create over 400 designs every second, so it used model distillation to cut the number of inference steps to a fraction of the original, and built a custom inference engine for the planning stage. The company says that made Craft an order of magnitude faster than its first internal version. It does not publish latency figures, parameter counts or the size of either stage.

Trained on templates rather than photographs

Most image models learn mainly from photographs. Canva says Craft's training puts much more weight on design: its data includes millions of human-made templates from Canva's own content library, including posters, social posts, presentations, infographics, menus and typography-heavy layouts. The argument is that a layered template teaches a model how hierarchy, alignment and type fit together in a way a flat finished image does not.

After pre-training, Canva says it tuned the model with supervised fine-tuning and reinforcement learning using feedback from its professional designers. For internal evaluation, designers and dozens of researchers compared more than 25,000 pairs of images against leading models, and Canva says that when automated scores and human judgments disagreed, it sided with the designer.

There is a practical payoff inside Canva's own product. Because Craft plans images as compositions of separate elements, Canva says its outputs suit Magic Layers, the feature that converts a flat image into editable design layers. The post mentions JSON bounding boxes as the mechanism that makes this conversion faster and more accurate. In other words, a generated poster can become an editable Canva design rather than a single picture.

The post does not say whether contributors to the template library were told their work would train an image model, or whether they are paid for it. Licensing of training data has become a live issue across the industry, and it is one Canva's announcement leaves unaddressed.

The ranking claims, and who made them

Canva's headline claim is that Craft makes Canva Research "the world's second most advanced AI lab for image generation," measured by DesignArena, a crowdsourced site where users vote between outputs from different models. According to Canva, Craft debuted fourth overall among image generation models on DesignArena, ranked first in the People and Portraits category out of 87 models and in Landscapes and Nature out of 85, second in Abstract and Patterns, and fourth in Products.

Those are Canva's descriptions of a third-party leaderboard, and arena rankings move as new models arrive and votes accumulate. AI Lookout could not independently confirm the positions from DesignArena's public pages at the time of writing. The "second lab" framing also depends on counting each company once, so a lab with several highly ranked models counts the same as one with a single entry.

Design quality is hard to score. AI Lookout recently covered AtelierBench, which found no AI model reaching the professional design bar, a reminder that crowd votes on attractive images are not the same thing as a layout a working designer would ship. Canva itself says the first version still needs work on fine print and technical drawings.

Where you can use Craft

For ordinary users, Craft lives in Dream Lab, which Canva says is in the bottom left panel of the editor. Canva lists the use cases it is aiming at: posters, social posts, flyers, presentation visuals and ads; lifestyle and editorial-style photography; flat, vector, painterly and hand-drawn illustration; portraits and groups; landscapes and interiors; and conceptual compositions.

For developers, Craft is available through Canva's public API, which Canva pitches to marketing platforms, e-commerce sites, ad builders and print services. Canva also says Craft will connect to AI assistants through its Canva MCP server, but that part is marked coming soon. TechCrunch reported that the plan is to let chatbots such as ChatGPT and Gemini generate images through that route, so an agent could make a launch graphic as one step in a larger task.

Canva has not published API pricing for Craft in its announcement. The post says Craft-fast and Craft-ultrafast are designed to deliver high-quality images at a fraction of the cost of leading frontier models, which suggests price will be part of the pitch once those variants ship.

Safety controls fall under Canva Shield, the company's umbrella AI-safety programme. Canva says that includes input and output moderation, privacy controls and reporting tools, and that models are evaluated for brand safety and cultural context before release. It does not describe watermarking or provenance labels for Craft images.

Why Canva wants its own model

The business logic is easy to see. Canva has used a range of outside image models inside its product, and TechCrunch noted that an in-house model is likely to save on API costs. Canva itself says owning the model gives it more control over quality and speed. At Canva's scale, the cost of every generated image adds up quickly.

TechCrunch also reported that Canva was on track for about $4 billion in revenue last year, and that analysts have cut their estimates of its valuation from $42 billion to $35 billion after it lowered its 2026 growth forecast. A model it owns, and can sell through an API, gives Canva a new product line as well as lower costs.

The release lands in a crowded month for image models. AI Lookout has covered Black Forest Labs' FLUX.3 Image Klein open weights, Krea 3 with 2K edits in under eight seconds and OpenMOSS's MOSS-Image 5B for long-text layouts. Legible text and layout control are where several of these models now compete, and Canva is betting that its template library gives it an edge there.

The pressure on stock imagery is part of the same story. AI Lookout reported this week that Getty Images filed for Chapter 11 and cited generative AI pressure on iStock. Canva, which sells both templates and stock media, now has its own generator sitting next to both.

What is still unproven

Nearly all of the performance evidence comes from Canva. The 25,000-pair internal comparison is not published, the DesignArena positions are described by Canva rather than shown in a dated snapshot, and there is no model card, technical report or list of the models Craft was compared against.

Canva also does not say how large Craft is, what hardware it runs on, how long a typical image takes, or what the API costs. The claim of being an order of magnitude faster is relative to Canva's own first internal version, not to rival models. Users will be able to test speed and quality for themselves in Dream Lab, but teams choosing an API will want numbers.

Finally, the training-data question matters more for a company like Canva than for most. Its template library is built partly by outside creators, and the post does not explain how that work was licensed for model training or whether creators can opt out.

What to watch next

Watch for the release dates and prices of Craft-fast and Craft-ultrafast, for the Canva MCP integration that would let ChatGPT, Gemini and other agents call Craft directly, and for any technical report or model card. Independent comparisons on DesignArena and other leaderboards over the next few weeks will show whether the launch rankings hold. Canva also promises reference-image and more advanced editing features in later releases.