Ente Photos

Ente Photos is an end-to-end-encrypted photo storage product developed by Ente, a company building privacy-preserving consumer infrastructure. The codebase is fully open source under github.com/ente-io/ente and covers both client apps (mobile, desktop, web) and server, with self-hosting supported. The client is built primarily in Flutter; server-side storage uses encrypted objects keyed per file, with the user holding the master key.

Because Ente cannot read user photos by design, all machine-learning workloads run on the user's device. Magic Search, the natural-language image search feature, uses Apple's MobileCLIP model, which is an efficiency-oriented iteration of OpenAI's CLIP architecture. The MobileCLIP image encoder runs locally over each photo to produce an embedding; at query time the text encoder runs over the user's prompt and the two embeddings are compared inside the same vector space. Face recognition runs YOLOv5-based face detection (the FlutterFace demo repo, github.com/laurenspriem/flutterface, documents the implementation Ente uses to explore on-device face recognition) followed by MobileFaceNet for face embeddings.

Image embeddings and face embeddings are compressed losslessly and encrypted with the per-file key before being uploaded to object storage. Clustering and person-grouping happen locally; the resulting "person entities" are themselves encrypted before leaving the device. Once derived, the ML index is synced end-to-end encrypted across the user's devices, so indexing done on a high-end laptop can be reused on a phone without re-running the model from scratch.

Mozilla selected Ente Photos for the first cohort of the Mozilla Builders Accelerator, announced September 23, 2024, with awards of up to $100K plus mentorship. The grant is non-dilutive. The cohort was scoped around Local AI, AI that runs on the user's hardware rather than in a vendor cloud, and Mozilla Builders rules require participating projects to release their work under open-source licenses. The grant sits at the retrieval-memory and runtime layers of the open stack: it pushes back on the default that consumer photo libraries are vendor-readable training data.