Google Photos Magic Editor: GenAI Under the Hood of a Billion-User App
summary
Google Photos integrates traditional CV and large generative models to power features like Magic Eraser and post‑capture portrait segmentation, serving 1.5 billion monthly users and doing hundreds of millions of edits per month. Engineering uses a shared C++ client library, on‑device inference via TensorFlow Lite/LightRT and EdgeTPU when possible, but moves heavyweight models to servers (hundreds of MB) for quality, exposing server capacity, latency, and testing tradeoffs; models work well for many shots but fail on fine hair, prompting evals, distillation, and UI constraints.
AI-written summary · about
chapters (16)
- 0:00Intro
- 1:28Google Photos Computational Photography Team
- 4:46Post Capture Segmentation
- 7:21Magic Eraser
- 8:12ML Learnings
- 10:02What happened in 2022
- 10:44Why GenAI
- 12:04Challenges
- 12:56Client Server
- 14:10Problem Space
- 16:06Reducing Ambiguity
- 17:24Trust and Safety
- 18:07Working with AI
- 19:07Whats next
- 19:33Personal view
- 20:13Conclusion
related talks
37mEnhancing Media Experiences in the Age of AI: A Practical Guide
25mPersonalizing Your Images with AI-Powered Features
30mImproving Photomath’s math problem scanning with deep learning
01h 18mComputer Vision after the Victory of Data
36mApps That See 👀 Practical AI Vision for Developers
38mHow Artificial Intelligence has changed your life
33mSee More, Be More
41mKEYNOTE: Image processing, AI, and autonomous systems