ALT A table comparing in the cloud vs on device approaches: Feasibility: More size & computing power vs Limited size & computing power Cost: Cost per API usage vs No cost per API usage Latency: Slower because of API call vs Slower because of limited power Availability: Device online, server working vs All time App size: Smaller vs Generally bigger Privacy: Generally harder to manage vs Generally easier to manage Model updates: Any time vs Depends on implementation More suitable for: Models won't fit/work in a device, infrequent usage, latency is acceptable vs Smaller/faster models, frequent usage, performance is important

ALT A table comparing on-device approaches: In the OS vs In the main app vs Dynamic delivery vs From a server Cost: No download cost vs No download cost vs No download cost vs Download cost Available: Where Google wants vs All time vs When installation succeeds vs When download succeeds Size for users who can't use: Small or no increase, depends on implementation vs Increases vs Doesn't increase (on-demand) vs Doesn't increase Updates: When Google wants vs With app updates vs With app updates vs Any time More suitable: No customisation vs All users use feature, same model for all vs Some users use feature, same/few models for all vs Many different models, smaller models Examples: No usage vs Prototypes vs Camera vs Typing