Canva has released Canva Craft, the first image generation model built and trained from scratch by Canva Research, its in-house AI lab. The company announced it on October 22 in a newsroom post, which landed early on October 23 in Manila. Craft is available to all Canva users globally through Dream Lab and to developers through Canva’s public API.
Bloggers, marketers and small teams already make pins, thumbnails and social graphics in Canva every day. Craft is the first image model Canva Research has built and trained for that work, and Canva says its training puts much greater emphasis on design than most image models do.
How the Two-Stage Model Works
Canva says conventional image models generate pixels directly from a prompt, and that Craft splits the job in two.
- A planning model reads the prompt and writes a detailed plan of the composition. According to Canva, that plan covers which elements appear, their spatial relationships, relative scale, visual hierarchy and any text that needs to appear in the image.
- A separate rendering model then draws the image from that plan at up to 2K resolution, in more than 30 aspect ratios.
Canva says planning first lets Craft produce complex layouts with “legible text, precise object placement and consistent visual relationships,” and allows targeted changes to one element without altering everything else. Those are Canva’s claims. The company also says the structure suits Magic Layers, its tool that turns a flat image into editable design elements. Search Engine Watch notes that Magic Layers can rebuild text in an image as editable text boxes, so a heading’s wording or font can be changed without redrawing the picture.
Prompts can be written in any language, and Canva says Craft can render text in 55 languages. You can choose one of the preset aspect ratios or enter your own.
What Canva Trained It On
Canva says Craft’s training data includes “millions of human-made templates” from its content library, including posters, social posts, presentations, infographics, menus and typography-heavy layouts. The company then fine-tuned the model with supervised fine-tuning and reinforcement learning, using feedback from its professional designers. Canva says those designers and dozens of researchers assessed more than 25,000 pairs of images in head-to-head comparisons with leading models, and that when automated evaluations and human judgments differed, it sided with the designer. Canva hasn’t published that comparison set.
TechCrunch reports that researchers from Leonardo.ai, the image generation company Canva acquired in 2024, contributed to the model. TechCrunch also notes that Canva has used several outside image models in its platform and suggests an in-house model will likely help Canva save on API costs.
Where Craft Ranks
Canva says Craft debuted as the fourth most advanced image generation model on DesignArena, a crowdsourced leaderboard where users compare AI outputs, and that this makes Canva Research the second most advanced lab for image generation. At launch, The Verge reported that the three models ahead of Craft were OpenAI’s GPT-Image-2.5 Sunburst, GPT-Image-2.5 Flare and GPT-Image-2. By the morning of October 23 in Manila, a different model held fourth place on DesignArena’s live image leaderboard.
Canva also says Craft ranks first in People and Portraits and in Landscapes and Nature, second in Abstract and Patterns, and fourth for Products on DesignArena. Those subcategory placements are Canva’s figures. Leaderboard positions move as votes come in, so treat all of these rankings as a snapshot from launch day.
The Text Claim Has a Caveat
Legible text is the headline pitch. OrcaRouter, an AI model routing service, points out that Canva’s own API reference recommends a different model for text-heavy images. Canva’s API reference lists three image models. It recommends Craft for infographics, posters, slides, portraits, landscapes, products and marketing materials, and it recommends Z Image Turbo “if the image includes text.” The reference says those strengths are “based on Canva’s evaluations” and tells developers to test their own prompts before pinning a model.
OrcaRouter calls this a hedge, since Craft’s own list includes posters and infographics. The practical point for anyone making text-heavy graphics is to check the spelling and placement on your own designs before you rely on it.
How to Try It and What It Costs
Craft is live in Dream Lab. The Verge says you can reach it from the three-dot option in Canva’s left-hand menu, and Canva’s post points to Dream Lab in the bottom left panel. Canva hasn’t published a per-image credit cost for Craft. Its developer announcement says starting a job uses your AI credits, the same as its other models, so each Craft image counts against your plan’s AI allowance.
For developers, Canva’s developer forum says the model value is craft_v1_2k on the image generation endpoint. A job takes roughly 50 to 60 seconds because Craft plans before it renders, and the download link stays valid for 24 hours. Canva’s API reference says the generation APIs are a preview, which means breaking changes can happen, and public apps that use preview APIs won’t pass Canva’s review.
What’s Planned Next
- Canva MCP support is planned. Canva lists Craft on the Canva MCP Server as “coming soon,” which would let AI assistants and automated workflows create images ready to use in a design. TechCrunch reports Canva plans to support image generation in chatbots such as ChatGPT and Gemini through MCP. None of this is live yet.
- Canva says faster, cheaper variants called Craft-fast and Craft-ultrafast are on the way. It hasn’t given a release date.
- Canva says more advanced editing and reference image features will come in future releases, along with better handling of fine print and technical drawings.
Early Reactions
Press coverage arrived the same day. The Verge described Craft as a model that “competes directly with those from OpenAI, Microsoft, and Google.” Search Engine Watch focused on the link between Craft’s planning step and Magic Layers. Community discussion beyond launch posts is still thin.
On X, DesignArena said Craft “wins most often on photorealistic product and object shots.” Product newsletter writer Lenny Rachitsky quoted Canva’s launch post with “Canva the frontier image model,” and Dhaval Makwana wrote that Craft “thinks through the layout first, decides where the text and important elements belong, then generates the result.” OrcaRouter notes that no outlet had published side-by-side test outputs when it wrote its analysis.
JamoraquAI Take
Most AI image tools are good at mood and weak at the dull parts of design, like spelling a headline right, putting a logo where you asked and changing one word without repainting the whole poster. Those dull parts are what bloggers and small marketers need from a pin or a thumbnail. Canva’s plan first, draw second approach is aimed squarely at them, and it ships inside an app many of us already open every morning.
The caveats are real. The rankings are a launch-day snapshot, the quality claims are Canva’s, Canva’s own API reference still recommends another model for images with text, and every image uses your AI credits. The MCP and assistant connections aren’t live yet either.
The test costs you a few credits. Run your next Pinterest pin or YouTube thumbnail prompt through Dream Lab, then check the text letter by letter. If it comes out right the first time, that saves a trip to another tool. If it doesn’t, Magic Layers may let you fix the words without starting over.
Sources
- Canva Newsroom, “Introducing Canva Craft” (October 22, 2026)
- Canva Developers Community, “Canva Craft is now available for image generation” (Connect API) (October 22, 2026)
- Canva Developers, Create image generation job, REST API reference (checked October 23, 2026)
- The Verge, Emma Roth, “Canva’s first AI image model plans designs before generating them” (October 22, 2026)
- Search Engine Watch, Martin Ashmere, “Canva Craft launches with 2K images and more than 30 aspect ratios” (October 22, 2026)
- TechCrunch, Ivan Mehta, “Canva releases its first image generation model” (October 22, 2026)
- OrcaRouter, “Canva Craft Debuts Fourth on Design Arena, but Canva’s Own API Sends Text Elsewhere” (October 2026)
- DesignArena on X (October 22, 2026)
