Canva Launches In‑House Image Generation Model 'Craft' to Power Design Workflows

Canva has unveiled its first proprietary image generation model, named Craft, built by its research team to rival offerings from Google, OpenAI, ByteDance and Meta.

The model operates in two stages: a planning phase that maps spatial relationships, scale, visual hierarchy and text placement, followed by a rendering phase that produces up to 2K resolution images across 30 aspect ratios. This architecture aims to deliver accurate object details, legible text and the ability to edit specific regions without affecting the rest of the canvas.

Training relied on Canva’s own library of human‑designed templates — posters, social posts, presentations, infographics, menus and typography‑rich layouts — rather than photographs. Outputs can be turned into layered designs via Canva Magic, the platform’s image‑to‑design transformer, and the model is accessible through the Dream Lab tab, the Canva API and an upcoming MCP integration for third‑party chatbots such as ChatGPT and Gemini.

Why it matters for GPU / AI infrastructure

• In‑house model reduces reliance on external API calls, lowering inference cost for a platform serving millions of users.
• Training on design‑centric data creates a niche workload that benefits from specialized GPU pipelines (e.g., high‑resolution rendering, text‑aware attention).
• Faster iteration cycles enable Canva to optimize model size (Craftfast, Craft‑ultrafast) for edge deployment, a trend relevant to any AI‑GPU cloud provider.

Canva’s acquisition of Leonardo.ai last year contributed research talent to Craft, and the company plans cost‑efficient variants for broader rollout. While Canva’s revenue trajectory remains strong, the new model underscores a shift toward vertical‑specific generative AI that can be hosted on dedicated GPU clusters.