Photo Scrubber blurs faces in photos

Summary

Simon Willison shared Photo Scrubber, an experimental browser-based tool. The tool uses Google's MediaPipe library, compiled to WebAssembly, along with the BlazeFace model to locally detect and blur faces in photos, while also stripping metadata. Willison noted that GPT-6 Astra generated this tool when he wanted to share protest photos without exposing identifiable bystanders.

Why it matters

The tool addresses a real and common privacy gap. People often share event photos where strangers' faces appear unintentionally. Most existing solutions require uploading images to cloud services. Photo Scrubber, however, runs entirely in the browser via WebAssembly, keeping images on the device. This is significant in an era of increasing surveillance and data collection. The tool also serves as a concrete example of large language models (GPT-6 Astra) generating practical, single-purpose tools from natural language prompts.

Highlights

• Runs entirely client-side using MediaPipe C++ compiled to WebAssembly — no data exchange with a server
• Generated by GPT-6 Astra as an 'experimental' tool from a single-line use case
• Combines face blurring (BlazeFace) and metadata stripping in a single step

Critical take

The article is essentially a two-sentence tool announcement. There are no accuracy comparisons. Edge cases such as partially visible faces, small faces in crowds, or side angles are not discussed. The other half of the promise, 'metadata stripping,' is also not explained. Labeling the tool as 'experimental' carries a heavy burden here. Additionally, the fact that the tool was generated by AI without visible human review raises legitimate questions about robustness before trusting it with a real photo set.