Smart glasses are a promising platform for vision-language assistants, but private split inference is expensive. The arXiv paper introducing SAGE notes that protecting visual representations before transmission does not remove the cost of on-device vision encoding, which remains a major bottleneck.
The authors propose SAGE as an efficient approach for protected vision-language inference on smart glasses. While the abstract does not detail the method, it highlights a key inefficiency in current private inference pipelines: the overhead of encoding visual inputs on the device itself.
If SAGE delivers on its promise, it could bring practical private vision-language assistance to smart glasses by reducing the computational burden that has so far limited such systems.