Canva Launches Craft, a Two‑Stage AI Image Model That Plans Designs Before Rendering

Canva has unveiled Craft, an internally developed image generation model that competes directly with offerings from OpenAI, Microsoft, and Google. The system employs a two‑stage architecture: a planning phase that maps out layout, text placement, and object relationships before a rendering stage produces the final pixel output.

Craft supports resolutions up to 2K and more than 30 aspect ratios, enabling everything from social‑media posts to detailed environmental scenes. According to Canva, the model was trained on millions of human‑created templates from its content library, giving it a strong grasp of design conventions and legible typography.

On the crowdsourced Design Arena leaderboard, Craft currently ranks fourth, trailing only the latest GPT‑Image variants. The release also includes robust input and output moderation, privacy controls, and an open Model Context Protocol for future AI‑agent integration, with faster Craft‑fast and Craft‑ultrafast variants planned.

Why it matters for GPU / AI infrastructure

High‑resolution, multi‑ratio generation at scale drives demand for low‑latency inference on powerful GPUs. As platforms like Canva embed such models into everyday design workflows, cloud providers must offer flexible, high‑throughput GPU capacity to keep rendering times acceptable for millions of concurrent users.