Canva Craft Is Canva's First Image Model, Built From Scratch

Canva just stepped into the model race. On October 22, the design company released Craft, the first image-generation model built and trained from scratch by Canva Research, and the lab's first public model release. The Canva Craft model goes straight at image models from Google, OpenAI, ByteDance, and Meta, which is a bold move for a company the world knows as an app, not an AI lab.

Until now, Canva rented its image generation the way most apps do. Dream Lab, Canva's AI image tool, launched in October 2024 running on Phoenix, the model from Leonardo.ai, a media-generation platform Canva acquired that same year. Craft changes the arrangement. Researchers from Leonardo.ai contributed to the new model, but this time the weights, the training, and the roadmap belong to Canva.

What the Canva Craft model actually does

Craft works as a two-stage pipeline. A planning model first reads your prompt and writes a detailed composition plan: spatial relationships, scale, visual hierarchy, what text goes where. Then a rendering model turns that plan into the finished image. Splitting planning from rendering is Canva's answer to the classic AI image problem of prompts that produce pretty pictures with the wrong layout.

The output specs are aimed squarely at working designers. Images render at up to 2K resolution in 30 aspect ratios, plus custom sizes, so one prompt can cover a poster, a story, and a presentation slide. Text renders legibly in 55 languages, which matters more than it sounds: garbled text has been the giveaway of AI-generated design work for years. Individual elements can be edited precisely without redrawing the rest of the image, and generated images can be converted into layered, editable designs through Canva Magic, the company's image-to-design transformer.

Trained on designs, not photos

Here is the unusual part. Instead of training on photographs scraped from the web, Canva trained Craft on designs: human-designed templates from its own library, including posters, social posts, presentations, infographics, menus, and typography-heavy layouts.

The logic is sound. Most image models learn what the world looks like; Craft learned what good design looks like. If your job is making a flyer rather than a photorealistic landscape, a model that understands hierarchy, whitespace, and type pairing is more useful than one that knows what a sunset looks like. It also sidesteps some of the messiest training-data questions in the industry, since the templates come from Canva's own library rather than the open web.

Fourth on debut, and Canva is fine with that

On DesignArena's public leaderboard, where human voters rank image models blind, Craft debuted in fourth place. The three models ahead of it are all from OpenAI: GPT-Image-2.5 Sunburst, GPT-Image-2.5 Flare, and GPT Image 2. Canva is framing this as making it the second-best AI image lab in the world after OpenAI, which is a generous reading of fourth place but not an absurd one. For a first release from a lab that did not exist in the public eye until this week, fourth is a genuine result.

The more interesting detail is what comes next. Canva plans two cheaper variants, Craftfast and Craft-ultrafast, aimed at speed and cost rather than maximum quality. That tells you what this is really about.

The cost play behind the launch

Owning the model stack trims Canva's third-party API spend and gives the company control over quality, speed, and the feature roadmap. Every image generated in Dream Lab previously meant paying someone else's model; now the margin stays in-house. It is the same vertical-integration logic driving every big app company right now, applied to design: if image generation is core to your product, renting it forever is a tax you eventually stop wanting to pay.

There is also a platform angle. Craft is live in Dream Lab for Canva users, available through Canva's public API for developers, and coming soon via the Canva MCP Server so AI agents can call it directly. Canva says it plans to put image generation inside third-party chatbots like ChatGPT and Gemini through MCP, which would make Craft infrastructure, not just a feature. Media generation is having a crowded month: VideoGen's Model 1 is promising ten-minute AI videos at half the cost, and diffusion techniques keep pushing generation speed higher across the industry.

The honest read: Craft is not the best image model in the world, and Canva knows it. It does not need to be. It needs to be good enough that 200 million Canva users never think about which model is rendering their poster, while Canva stops paying rent on every generation. On that measure, fourth place on debut is a strong start.