People who make posters, social posts, or presentation slides often run into the same problem with AI-powered image generators: the pictures look sharp, but the words come out garbled or the layout falls apart once you try to place a headline next to a product shot.
This is the very reasin why designed tools keep looking for better ways to turn a short description into something ready to use.
Canva has long been a major player in providing a graphic design platform to create visual content for presentations, websites, and other digital products.
And this time, it has introduced 'Canva Craft,' the first image generation model built and trained from scratch by its own research group.
The model is now open to every Canva user through Dream Lab, and developers can also call it through the company's public API. Canva Research, introduced a day earlier, is the in-house lab behind it.
It has described Craft as a frontier model made specifically for design work rather than general photography.
Most image models take a prompt and produce pixels in one step.
Craft splits the process.
A planning stage first reads the request and writes out the composition: which elements should appear, where they sit relative to one another, how large each should be, the visual hierarchy, and any text that must be readable. A separate rendering stage then turns that plan into a finished image at up to 2K resolution across more than 30 aspect ratios, or a custom size.
The company says the separation helps with precise object placement and legible lettering, and it also makes it easier to change one part of an image without rewriting the rest.
The training data reflects the same focus.
Instead of relying mainly on photographs scraped from the open web, Craft was trained on millions of human-made templates already stored in Canva's library.
Those templates include posters, social posts, presentations, infographics, menus, and other typography-heavy layouts. Researchers who joined after Canva acquired Leonardo.ai in 2024 also worked on the project.
Internal tests involved professional designers comparing more than 25,000 image pairs against other models; when automated scores and human judgments differed, the designers’ view took priority.
On DesignArena, a crowdsourced leaderboard that ranks models by blind preference votes, Craft entered in 4th place overall with an Elo score of 1340. The three models above it are from OpenAI.
Canva therefore ranks its research lab second among AI labs on that particular board.
In subcategory rankings the model placed first for people and portraits out of 87 models and first for landscapes and nature out of 85, and second for abstract patterns. It also supports text rendering in 55 languages while accepting prompts in any language.
To use the feature, Canva users can open Dream Lab from the bottom-left panel inside Canva and generate images with Craft. Outputs can be turned into layered, editable designs through Canva's Magic Layers feature.
The public API exposes the same model under the identifier craft_v1_2k. Canva has said it intends to add support for AI agents through the Model Context Protocol (MCP) and to release faster variants called Craft-fast and Craft-ultrafast later.
With Craft, Canva is entering a highly competitive market where its primary rival, Adobe, has set the industry standard with its Adobe Firefly engine.
While Firefly focuses on generating high-fidelity, commercial-safe assets for seasoned professionals inside Photoshop and Illustrator, Craft shifts the focus to immediate layout usability.
By training on millions of human-made templates rather than raw stock photography, Canva has optimized its model for quick marketing workflows, precise object placement, and multi-language text legibility that fits directly into everyday digital products.
