Canva Craft: Its First In-House Image Model Is Trained on Templates, Not Photos, and Plans the Layout Before It Draws
Craft is live for every Canva user in Dream Lab and through the Canva API, renders up to 2K in more than 30 aspect ratios, and writes text in 55 languages. Canva hasn't said what it costs in credits, and cheaper Craft-fast versions are still to come.
TL;DR
Canva released Canva Craft on 22 October 2026, its first image generation model built in-house by Canva Research, with help from the Leonardo.ai team it bought in 2024. Craft first plans a layout (spatial relationships, scale, hierarchy and text) and then renders images up to 2K in more than 30 aspect ratios, with text in 55 languages. It was trained on human-designed Canva templates rather than photographs and is available to all users through Dream Lab and to developers through the Canva API. Canva says it debuted fourth on the crowdsourced DesignArena leaderboard, but it has not said which plans include it or how many AI credits it uses.
Canva has relied on outside image models, including the Leonardo Phoenix model behind the original Dream Lab, since it started offering AI images. Craft is the first one it built itself. Its pitch is that it is a design model rather than a photo model: it is meant to make posters, social posts and menus with legible text and layouts you can edit, not photorealistic scenes.
What Craft does
- Two stages: a planning model maps the elements, their positions, relative scale, visual hierarchy and any text, then a separate renderer produces the image.
- Output: up to 2K resolution, more than 30 aspect-ratio presets plus custom ratios, and generated text in 55 languages. Prompts can be in any language.
- Editing: Canva says Craft can make precise edits that leave the rest of the image unchanged, and its layout data helps Magic Layers turn a generated image into separate editable elements.
- Safety: covered by Canva Shield, Canva's moderation and safety controls on inputs and outputs.
Where it ranks, and how far to trust that
| DesignArena category | Craft's position |
|---|---|
| Overall image models | 4th |
| People and portraits | 1st of 87 |
| Landscapes and nature | 1st of 85 |
| Abstract and patterns | 2nd |
| Products | 4th |
TechCrunch reported no standard benchmarks. Canva says designers and researchers compared more than 25,000 image pairs during development and gave designer judgment priority over automated scores. The rankings are a useful signal, but they come from Canva's own reporting of a voting leaderboard.
What Canva hasn't said
- Price: neither report says which Canva plans include Craft or how many AI credits a generation uses. Canva's AI tools draw on a shared monthly credit allowance, so check your plan's usage after a few runs.
- Training data terms: Canva says Craft learned from human-made templates in its library, but has not published how template creators are credited or paid for that use.
- Chatbot access: TechCrunch says Canva plans to offer Craft inside ChatGPT and Gemini through its MCP connector; TestingCatalog describes MCP support as coming soon. Treat it as not yet live.
- Cheaper models: Craft-fast and Craft-ultrafast are in development with no release date.
What this means if you choose image tools
- For marketing graphics with text on them, Craft is worth trying first if you already pay for Canva. Legible text and editable layouts are the jobs it was trained for.
- For photorealistic product shots or illustration, keep comparing it with dedicated image models. Its training on templates rather than photos points it at design, not photography.
- Developers can call it through the Canva API today, which makes it an option for automated social graphics.
- Watch for Canva's credit pricing. TechCrunch expects the in-house model to cut what Canva pays outside model vendors, but Canva hasn't said whether users will see that saving.
Sources
AI Industry Reporter
Priya covers model releases, industry announcements, and the gap between what labs claim and what independent evaluators actually find. She reads the primary source - the paper, the system card, the benchmark org's own statement - before writing a word.
More on AI Image Generators
Pruna's P-H3-Image Makes 8-Megapixel Images in 8 Steps, but Its License Excludes the US, EU, UK and South Korea
Priya Nair · 4 min
Pangram Image 1.0 and Image Flash 1.0: AI Image and Video Detection Leaves Preview, With Heatmaps and Video Uploads
Priya Nair · 4 min
OpenAI Stops Taking ChatGPT Ads From Rival Image and Audio Generators. What Adobe Said, and the Questions OpenAI Hasn't Answered
Priya Nair · 4 min