Skip to content

performance.md

Latest commit

 

History

History
20 lines (15 loc) · 919 Bytes

performance.md

File metadata and controls

20 lines (15 loc) · 919 Bytes
Raw

Performance

The v1.2 selector reduces scene scanning work and API payload size before changing model-side latency.

On the bundled 14.04-second, 351-frame development fixture, the selector inspected 31 frames (91.17% fewer than a full-frame scan). A final local comparison measured about 2.23 seconds for v1.2 versus 2.72 seconds for the legacy scan, a 1.22× speedup. These numbers describe one machine and codec, not a universal guarantee.

Reproduce the comparison:

python benchmarks/benchmark_frame_selector.py Examples/pizza.mp4

For cloud workloads, the main cost controls are max_frames, frames_per_minute, max_image_dimension, JPEG quality, audio enablement, and model selection. Frame analysis is concurrent up to max_workers; increase it only within provider rate limits. Metadata contains per-stage wall time and provider usage counts so applications can measure their own workloads.