Australia's Canva Launches Its First In-House AI Image Model, Craft: Plans the Layout, Then Renders Up to 2K with Text in 55 Languages

Canva on October 22 released Canva Craft, its first self-built image generation model: a planning model maps out the composition before a renderer generates images at up to 2K resolution across 30+ aspect ratios, trained on millions of human-designed templates rather than photographs. Live for all users in Dream Lab and via public API.
Why it matters
Australian design platform Canva released Canva Craft on October 22, its first AI image-generation model built from scratch by its in-house Canva Research lab. Until now Canva's AI image tools relied on external models (including Phoenix from Leonardo.ai, the Sydney startup it bought in 2024). Going in-house puts it in direct competition with image models from Google, OpenAI, ByteDance and Meta — and is widely seen as a move to cut third-party API costs.
Details
Craft's signature is a two-stage architecture: a planning model first turns the prompt into a detailed composition plan — which elements appear, where they sit, their scale, visual hierarchy and any text — and a separate rendering model then generates images at up to 2K resolution in more than 30 aspect ratios (or a custom size). Canva says this yields legible in-image text, precise placement and targeted edits to one element without redrawing the whole picture. Prompts can be written in any language; generated text is supported in 55 languages.
Unlike most image models trained on photographs, Craft was trained on millions of human-designed templates from Canva's library — posters, social posts, presentations, infographics, menus — so outputs can be converted into editable, layered designs via Magic Layers. Canva says Craft debuted fourth overall on DesignArena's public leaderboard, first in People & Portraits and Landscapes & Nature, and calls Canva Research the world's second-best image lab behind OpenAI — the company's own claim.
Availability: all users get it in the Dream Lab tab now; developers can call it through the public API (model craft_v1_2k). Researchers from Leonardo.ai contributed to the build, cheaper/faster Craftfast and Craft-ultrafast variants are coming, and ChatGPT/Gemini integration via MCP is planned. Context: external analysts previously reported Canva cut its 2026 growth forecast, lowering its valuation from $42 billion to $35 billion — an in-house model is a cost-control play.
Sources
Canva Developer Community official announcement (2026-10-22); SiliconReport (2026-10-22); madrobot.blog (2026-10-22, originally TechCrunch); testingcatalog.com (2026-10-23).
N環球 
💬 Comments
No comments yet. Be the first!