PS5CameraDriver 1.2.9

Version 1.2.9 — 14 October 2026

Adds native macOS installer checks on Apple Silicon and Intel, and observable Windows video
timing through ps5cam-ctl latency. Picture settings and calibration are preserved on upgrade.
This release does not add a 1080p30 mode or change the video-quality algorithms.

The macOS installer uses binary-only dependencies with exact versions and SHA-256 hashes,
including the complete Python 3.9 dependency graph. Both native CI jobs install the assembled
ZIP, check the libusb architecture and launchd service, and uninstall it. Firmware fixtures
are synthetic; Sony's firmware is not included in the release.

Windows Device MFT keeps the newest undelivered frame instead of a three-frame FIFO. Replaced
frames are counted and marked discontinuous. Device timestamps and timing metadata follow the
actual image returned by the one-frame GPU readback pipeline, including after flush/recovery.

Close other camera applications and measure 360 frames after 30 warm-up frames:

& "$env:ProgramFiles\PS5Camera\ps5cam-ctl.exe" latency 360 0 latency.csv

Use 360 40 slow-consumer.csv to simulate a slow application. Reports contain mean, p50, p95,
maximum and coverage for available metrics; the CSV contains timings and counters, without images.
Unavailable stages remain explicitly unavailable. The measurement ends at the local Media
Foundation reader; preview/display, encoding and network latency require separate measurements.

In short 1080p60 bokeh tests on an RTX 3060 Ti / Ryzen 5 5600, the local 1.2.9 candidate delivered
59.83 fps with a mean capture-source-to-reader time of 34.70 ms (1.2.8: 34.12 ms). A 40 ms pause
between reads produced about 109 ms in both versions. The new telemetry located 75.55 ms of
that delay after the driver had delivered the sample, while the driver's ready queue averaged
0.029 ms and GPU compute 6.54 ms. The newest-frame policy is covered by a production Device MFT
regression; this reader test did not demonstrate lower total latency. The candidate test also ran
alongside Sandbox, so the small difference between the short series is not an isolated benchmark.