Canva releases Canva Craft, its first in-house AI image generation model
Canva released Canva Craft, its first in-house AI image generation model, developed by its research team to compete with models from Google, OpenAI, ByteDance and Meta. The two-stage system uses a planner that defines composition, layout and exact text before a renderer produces images at up to 2K resolution across more than 30 aspect ratios. Trained on human-designed Canva templates rather than photographs, it is available to all users via the Dream Lab tab and the Canva API, with third-party chatbot support via MCP planned.
IllustrationEditorial responsibility
- No named human review is recorded for this page.
- Reports are grouped by semantic similarity and deterministic rules. Language models may assist titles, summaries, translation and cross-source analysis; the page reads the event directly, while its address stays stable when the title changes.
- Summary covers the current reports
Cross-source coverage
Reporting timeline
Canva Launches Canva Craft, Its First In-House AI Image Model
Canva has released Canva Craft, its first self-developed AI image generation model. The model uses a two-stage architecture that plans the design before generating the image, supports resolutions up to 2K, and handles more than 30 aspect ratios. It was trained on millions of human-made templates from Canva's platform. Canva says the model currently ranks fourth on the Design Arena leaderboard. Canva Craft is now available to all users worldwide. The company plans to release Craft-fast and Craft-ultrafast versions in the future and to support connecting the model to AI agents through MCP.
Read sourceCanva Debuts In-House Image Model, Ranks Second on DesignArena Behind OpenAI
Canva has entered the frontier AI model race with its first in-house image model, which placed as the second-best lab in image generation on DesignArena, with only OpenAI ahead of it. The system, called Canva Craft, uses a two-model architecture: a planner that reads the prompt and writes out the composition before any pixels exist, specifying which elements appear, where each sits, its size, what carries emphasis, and the exact words to display. A second renderer model then paints that plan at up to 2K resolution across more than 30 aspect ratios. This makes layout an explicit intermediate object rather than something the model must resolve in a single pass, and spelling is settled before rendering begins. The plan also aids revision, since changing one element alters only that element, leaving adjacent elements such as a product shot untouched when a subhead is fixed.
Read sourceCanva's Craft model plans posters like a designer before rendering pixels
Canva has introduced a new image model called Craft that plans a poster's composition before rendering any pixels, addressing a common failure where image models misspell text such as writing SALE as SLAE. According to the post, most image models go straight from prompt to pixels in one pass, treating words as texture. Craft instead uses a planner that specifies which elements appear, where they sit, their size, and the exact words, followed by a renderer that paints the plan at up to 2K. Text becomes placed elements, improving legibility and spelling, and changing one element leaves others intact. Training data leans toward design files with layers, hierarchy, and typography, drawn from 54 billion designs made on the platform. Designers and researchers compared 25,000+ image pairs against leading models, with the team siding with designers over automated judges. Canva says owning the model controls cost per image, relevant given 75M+ monthly AI tool users and a prior paused rollout. The research team tripled in a year, and on Artificial Analysis's Image Arena the model trails only OpenAI, ahead of Google and Microsoft.
Read sourceShow 2 older updatesHide older updates
Canva Launches First In-House Image Generation Model, Canva Craft
Canva released its first image generation model, Canva Craft, developed by its research team to compete with models from Google, OpenAI, ByteDance, and Meta. The model works in two stages: first interpreting prompts to plan spatial relationships, scale, visual hierarchy, and text generation, then producing images up to 2K resolution in up to 30 aspect ratios. Canva says this enables accurate object details, legible text, and precise edits. The model was trained on human-designed templates from Canva's library rather than photographs, allowing generated images to be converted into layered designs via Canva Magic. Canva Craft is available to all users through the Dream Lab tab, the Canva API, and planned third-party chatbot support via MCP. Canva says the model reduces API costs and improves quality and speed control. Researchers from Leonardo.ai, acquired in 2024, contributed. Two cost-efficient versions, Craftfast and Craft-ultrafast, are planned. Canva cut its 2026 growth forecast, with its valuation reportedly falling from $42 billion to $35 billion.
Canva launches Canva Craft, its first in-house image generation model
Canva has released its first image generation model, Canva Craft, developed by its research team to compete with models from Google, OpenAI, ByteDance, and Meta. The model works in two stages: first interpreting prompts and planning spatial relationships, scale, visual hierarchy, and text generation; then generating images up to 2K resolution in up to 30 aspect ratios. Canva says this approach produces accurate object details and legible text, and allows precise edits without altering other parts of an image. The model was trained on human-designed Canva templates rather than photographs, so outputs can be converted into layered designs using Canva Magic. Canva Craft is available to all users via the Dream Lab tab, through the Canva API, and soon in third-party chatbots via Canva's MCP. Canva says the model will reduce API costs and improve quality and speed control. Researchers from Leonardo.ai, acquired in 2024, contributed. Cost-efficient versions, Craftfast and Craft-ultrafast, are planned. Canva was on track for $4 billion in revenue last year but cut its 2026 growth forecast, with its valuation reportedly falling from $42 billion to $35 billion.