How we built AI face cropping for Images

Cloudflare details the general availability of AI face cropping in Images, a feature that automatically detects faces and crops images using `gravity=face` and a `zoom` parameter. The post explains how RetinaFace (a CNN) powers detection, outlines the bounding-box aggregation and scaling logic, and describes performance choices like a 1024×1024 inference cap. It also covers the migration from a Rust/CPU-based prototype to a GPU-backed Workers AI implementation to mitigate memory pressure, reduce instance memory to ~150 MiB, and improve reliability; the feature is already applied to author images on the Cloudflare blog and will be followed by additional AI image tools like background removal and generative upscale.