// the find
google-ai-edge/mediapipe
Cross-platform, customizable ML solutions for live and streaming media.
MediaPipe is Google's framework for building on-device ML pipelines — think hand tracking, pose estimation, face mesh, object detection — running in real-time on mobile, web, and desktop. It ships both a low-level graph/calculator framework (C++) for building custom pipelines and a higher-level Tasks API (Android, iOS, Python, web) for dropping in pre-built solutions. The target audience is mobile and embedded developers who need ML inference without a server round-trip.
The graph-based calculator architecture handles real-time streaming media well — backpressure, synchronization, and GPU/CPU handoff are built into the framework rather than bolted on. The Tasks API abstracts away the Bazel build system and protobuf graph configs for common use cases, which is the right call for 90% of users. Pre-trained models ship with the library, so hand tracking or pose estimation works out of the box without hunting for weights. The cross-platform story is genuine — the same pipeline logic runs on Android, iOS, and web via WASM.
Bazel is the build system and there's no escape from it if you touch the C++ layer — this alone rules out casual contributors and adds a steep onboarding tax on Linux/Mac, and it's essentially unusable on Windows without WSL. The legacy solutions (Holistic, Face Mesh, BlazePose) were deprecated in early 2023 and frozen 'as-is', but they still dominate tutorials and Stack Overflow answers, so newcomers constantly land on dead APIs. The README in the repo is literally a redirect stub — all real documentation lives on developers.google.com, which means docs go dark when Google reorganizes URLs. Custom model integration (Model Maker) is limited to fine-tuning existing architectures; bringing in an arbitrary ONNX or custom TFLite model still requires writing C++ calculators.