Optimize lower-tier devices

Prerequisites

• A failing or marginal row in the physical-device qualification matrix
• A reproducible package, camera configuration, and session duration
• Release-build measurements for frame rate, memory, and thermal state

Reduce the input workload

Start at 720p and 30 FPS. Increase resolution only when the target device matrix proves the higher setting remains stable. Avoid a second camera controller, per-frame image conversion, synchronous readback, and unnecessary preview duplication.

Reduce active processing

1. Keep one authored package active in each intended slot.
2. Disable background segmentation when no background feature is visible.
3. Remove beauty or makeup controls that are set to their neutral value.
4. Avoid starting processed-frame sampling unless the app consumes it.
5. Prefer one optimized package over chaining several packages with equivalent output.

The native runtime can selectively skip inference work and adjust recording workload on measured lower-tier Android devices. Host code should still avoid adding its own unbounded queues.

Use latest-frame backpressure

When a consumer cannot keep up, replace pending work with the newest frame. Never accumulate camera frames to preserve every callback; that increases latency and memory until the user sees an old preview.

Verify the optimized experience

Repeat preview, recording, streaming, background/foreground, camera switching, and navigation loops. An optimization is acceptable only when it preserves the required effect output and does not create lifecycle or media regressions.

Common errors

Next steps

• Qualify devices and effects
• Reduce application size
• Troubleshoot runtime problems

References

• Limits and performance
• Rendering pipeline