Canva launches Craft, its image model designed for design
Canva launches Craft, its first in-house image generation model, available in Dream Lab and via API. The system mainly targets design uses, with readable text, layered editing, and rendering up to 2K.
Canva has launched Craft, its first foundational image generation model designed and trained internally by Canva Research. Available in Dream Lab for all users and via a public API for developers, it targets images up to 2K with text in 55 languages. The message is clear: Canva wants to make image generation a production tool, not just a lab demo.
A model designed for layouts, not just pretty images
Craft stands out with a two-step approach: first planning the composition, then rendering the image. The model maps requested elements, their scale relationships, visual hierarchy, and text before a dedicated engine generates the visual. This separation aims to improve readability, placement accuracy, and targeted edits without redoing the entire scene.
In other words, Canva is not just trying to produce "pretty" images; the tool is calibrated for posters, presentations, infographics, menus, or social posts where text matters as much as the image. This is where Craft differs from models that excel in ambiance but get tangled when a headline must remain sharp. Users can also cut results into editable layers thanks to Magic Layers.
Why Canva emphasizes speed and cost
The company says it serves over a quarter of a billion people, with more than 400 designs created every second, explaining its obsession with inference costs and speed. According to Canva, model distillation reduced inference steps to a fraction of the initial volume, while an internal engine was optimized for the planning phase. Craft is now reportedly ten times faster than its first internal version.
Canva is also preparing two more economical variants, Craft-fast and Craft-ultrafast, to offer quality generation at a lower cost than competing state-of-the-art models. This strategy matters for the market: a high-performing but too expensive model quickly ends up in the museum of good intentions. Here, the key remains large-scale usage.
Solid results on DesignArena, according to Canva
Canva claims Craft debuted in fourth place on the DesignArena ranking, placing Canva Research as the site's second image lab. The model ranked first in the People & Portraits category among 87 models and Landscapes & Nature among 85, second in Abstract & Patterns, and fourth in Products. These positions provide a useful indication, even if based on the platform's measurement and Canva's statements.
The company also specifies that training relied on millions of human-designed templates rather than a corpus primarily focused on photography. After this pre-training, Canva added a supervision phase and reinforcement learning fed by feedback from its designers. Designers and dozens of researchers evaluated over 25,000 image pairs against competing models, favoring human judgment when it differed from automatic metrics.
What this changes for professional uses
Craft is available now in Dream Lab, accessible from Canva's lower-left panel, and via the public API for integration into other tools. Canva also announces upcoming MCP support to allow AI assistants and automated workflows to generate ready-to-use images within broader tasks. Advanced editing and reference image features are also planned later.
The positioning is clear: Canva wants to transform image generation into an integrated work component, not an isolated tool. With Canva Shield for moderation, privacy, and security controls, the company tries to maintain a usable framework in business. The classic question remains: will the tool be good enough to replace some back-and-forth between creatives and product teams, or just speed it up? Probably both, depending on needs and level of demand.
Key points
- Craft is Canva's first in-house image model.
- Available in Dream Lab and via public API.
- Rendering up to 2K and text in 55 languages.
- Canva announced 4th place on DesignArena in October 2026.
- Over 25,000 image pairs were used for evaluation.
- Craft-fast and Craft-ultrafast are announced soon.
By the numbers
- 2K — maximum resolution announced by Canva for Craft.
- 55 languages — supported generated text, according to Canva, in October 2026.
- 30+ formats — predefined or customizable aspect ratios.
- 25,000 image pairs — evaluated by designers and researchers, according to Canva Research.
- 400 designs per second — volume announced by Canva worldwide.