Canva Craft: Canva's First In-House Image Model
Shipped 22 October 2026, Canva Research's foundation model plans a layout before it renders — 2K output, more than 30 aspect ratios, images that break into editable layers; it debuted fourth on DesignArena and took first globally for people and landscapes
Tool Interface
Interactive tool will be available soon
Features
- ✓ Two-stage architecture: a planning model writes out the composition, element relationships, hierarchy and any text first, then a rendering model draws it
- ✓ 2K output across more than 30 aspect ratios, so results drop straight into posters, social posts and decks without another crop
- ✓ Trained for design rather than photography: the data leans on millions of human-made Canva templates, teaching how layout, hierarchy and typography fit together
- ✓ Outputs break into editable layers: paired with Magic Layers, a generated image becomes design elements you can change individually — something plain text-to-image cannot do
- ✓ Available where you work: inside Canva and through the public API, with Craft-fast and Craft-ultrafast announced as faster, cheaper variants
How to Use
- In Canva, open the three-dot menu on the left and choose Dream Lab — Craft lives there and can be invoked while creating a design
- Describe structure in the prompt: where the subject sits, what the text says, which element goes where — the planning stage reads all of it
- Run the result through Magic Layers to split it into editable layers, then rewrite copy, recolour and reposition layer by layer instead of regenerating
- For volume, wire up the public API and watch for the Craft-fast and Craft-ultrafast variants, which Canva says will cost materially less than leading frontier models
FAQ
What is Canva Craft?
The first foundation image generation model trained from scratch by Canva Research, Canva's in-house AI lab. It launched on 22 October 2026, is available to all Canva users and to developers through the public API, and Canva says it debuted as the fourth most advanced image generation model on DesignArena, making Canva Research the world's second most advanced lab in that category. Announcement: https://www.canva.com/newsroom/news/canva-craft
What does the two-stage architecture actually buy you?
Conventional models go straight from prompt to pixels. Craft first runs a planning model that writes out which elements appear, their spatial relationships, relative scale, visual hierarchy and any text, then a rendering model draws the image. Canva says this yields more accurate complex layouts — legible text, precise object placement, consistent visual relationships — and makes it possible to change one element without disturbing everything else. See https://www.canva.com/newsroom/news/canva-craft
Are those benchmark numbers real?
The rankings are Canva's own claims, sourced from DesignArena, which Canva describes as the largest crowdsourced assessor of AI design. The stated figures: fourth overall at debut; first globally for People & Portraits out of 87 models; first for Landscapes & Nature out of 85; second for Abstract & Patterns; fourth for Products. Treat them as vendor-reported and check DesignArena directly rather than taking them as settled fact.
Is it slow or expensive?
Canva published two engineering claims: model distillation cut the inference steps to a fraction of the original model, and a custom inference engine tuned for the planning stage made Craft an order of magnitude faster than its first internal version. Canva also announced Craft-fast and Craft-ultrafast variants positioned at a fraction of the cost of leading frontier models. Check Canva's own pages for current pricing.
Craft versus Midjourney or DALL·E — how do I choose?
Ask whether you need photographic feel or layout. Craft's training leans toward design — posters, social posts, decks, infographics, menus, typography-heavy layouts — so its edge is legible text, precise placement and editable layers. For photoreal texture or a distinctive art style, mature bitmap models still win. The genuinely new part is that output becomes an editable design file.
Can agents call it?
Watch the MCP angle. Canva's announcement says it plans to let AI agents plug into Canva Craft through the open Model Context Protocol. The specification is at https://modelcontextprotocol.io ; check Canva's own announcement for what has actually shipped: https://www.canva.com/newsroom/news/canva-craft