MediaPipe Web: Bringing cross-platform AI tech to the browser
summary
MediaPipe Web explains how C++ pipelines become browser-ready via WebAssembly (Emscripten) and TypeScript bindings, and how cross‑platform MediaPipe Tasks power image segmentation, face/hand/pose tracking and a new LLM Inference API that runs models like Gemma, Phi‑2, Falcon and StableLM. Performance claims: 640 tokens/sec input and ~40 tokens/sec streamed output on a 5‑year‑old MacBook Pro, Gemma 1.1 7B (≈8.6 GB) runs locally on a 2021 MacBook Pro, and a WebGPU image generator produces 512×512 images in under two seconds. Features include LoRA and an AI Edge Torch conversion path.
AI-written summary · about
chapters (12)
- 0:00Introduction to MediaPipe
- 0:21What is MediaPipe?
- 1:37MediaPipe Architecture
- 2:51MediaPipe Web and Meet
- 3:42Expanding the Ecosystem
- 4:40Third-Party Integration
- 5:17MediaPipe Tasks API
- 6:21Vision and Other Tasks
- 7:26LLM Inference in Browser
- 8:58LLM Quality and Adaptability
- 10:47Customization and GraphRunner
- 12:21Image Generation Demo
related talks
24mWeb Apps of the Future With Web AI
46mFrom ML to LLM: on-device AI in the browser
20mFrom ML to LLM: On-Device AI in the Browser
32mAdvanced on-device ML made easy
30mBeyond the Web of Today
39mThe AI PC and the Era of the Modern Web
31mFrom ML to LLM on device AI in the browser
19mOn Dynamic Multimedia Pipelines for Opening Browsers to New Formats