Divide the softISP statistics window into 16x12 zones and, when a white
point locus is tuned, vote only with zones whose implied gains lie within
zoneGateMargin of the locus - a large coloured light source can no longer
drag the white balance. Validated on-phone 2026-08-24: keyboard-scene A/B
shows a ~1700 K estimate shift caused solely by the gate (desk renders
neutral instead of blue); torch-toggle stability equal or better than r8
(settle 1.35-1.79 s, ON-edge repeatability +-0.011, CCM churn 21 vs 30
switches); AF regression clean. Evidence:
journal/camera/captures/2026-08-24-awb-zonegate-validation.md
Assisted-by: Claude:claude-fable-5
gitlab.freedesktop.org's on-demand archive endpoint 503/504s for hours
at a stretch: runs #25, #27, #30 and #31 all died in 'abuild checksum'
fetching the pinned libqmi tarball, outlasting the 3x30s in-run retry.
git clone is served from a different code path and holds up (verified
against all three repos while the archive endpoint was flaky), so clone
libqmi/ModemManager/libcamera (retried) and git-archive the pinned refs
into the aports, the same way the kernel and imsd tarballs are already
produced. The pins stay single-sourced in the APKBUILDs.
libcamera gains a checksum step: its committed sums were for the
fd.o-served tarball, and git-archive output is not byte-identical.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The softISP AWB is pure gray world, so panning swings ColourGains with
the frame average (white wall renders green when the frame is
warm-dominated, dark desk renders red when cool-dominated) and the CCM,
interpolated over a CT estimated from those gains, amplifies and
quantizes the swings.
Patch 0015 (ours, original work) adds optional per-sensor Awb tuning to
the simple IPA: a piecewise-linear locus of calibrated illuminant white
points in gain space that the gray-world estimate is clamped to
(clampMargin), a per-stats-frame EMA on the applied gains (damping)
with a fast-converge path for persistent illuminant changes, and CT
interpolated from the applied gains' position along the locus - on the
same scale as the tuning file's ccms table - instead of the generic
estimateCCT(). Tuning files without whitePoints keep the previous
behaviour exactly, so every other sensor is unaffected.
imx896.yaml gains the locus from the stock tuning blob's 10-point
per-illuminant AWB calibration (white points and their exactly
reciprocal gain triplets, cross-validated): the 7 on-locus points as
nodes, the 3 fluorescent points admitted via clampMargin 0.35. The
5000 K label is the blob's own proven anchor; 2856 K (A) is locked by
the fluorescent trio projecting onto the 2856-5000 chord at ~3960 K
(TL84); the rest are geometry-derived approximations.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Raw-RDI dark frames (phone face-down, minimum analogue gain) put the
pedestal at exactly 64 in 10-bit (mean 63.98, median 64, p1/p99 63/65)
= 4096 on the 16-bit scale. The softISP's dynamic estimate ran at 48;
the 16-code residual was amplified by the CCM into colour casts in
dark image regions.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adds imx896.yaml (CCMs for 2800/4000/5750 K) so the main camera gets
real colour instead of the uncalibrated fallback. The matrices are
numeric calibration data read out of the stock Android tuning blob -
measured spectral response of this sensor, not authored expression -
and are shipped by the project owner's decision, consistent with the
vendor-derived data this image already carries. Durable replacement is
to measure our own from a colour chart.
Note the file lists Adjust/Agc/Af as well as BlackLevel/Awb/Ccm: a
sensor-specific tuning file REPLACES the fallback's algorithm list, so
omitting them silently disables auto-exposure and autofocus.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Hardware finding from the first r4 run on the phone: the focus-loss
detector compared the settled scene against the maximum sharpness
recorded during the sweep, which was measured under a different AGC
state. On the FP6 that maximum read 4848088 and the same position two
stats frames later read 1995445, so the detector fired immediately and
the lens swept forever. Patch 0014 takes the reference after the lens
settles and requires four consecutive out-of-band frames.
Verified on the phone: r4 re-swept without end, r5 completes one sweep
and stays put.
The DW9784 VCM works and libcamera discovers it through the sensor's
ancillary link, but nothing in the software ISP ever moves it. Backport
the out-of-tree autofocus work onto v0.7.2:
- 0007-0009 are Vasiliy Doylov's focus control, contrast autofocus and
focus-loss detection from the softisp-playground branch;
- 0010 is Pavel Machek's Librem5-tested robustness work (centre-window,
brightness-normalised sharpness, two-phase sweep, settle skip),
squashed and adapted;
- 0011-0013 are ours: the lens write no longer sits inside the
no-frame-start-emitter branch, the lensless-camera paths are guarded,
and AfMode is advertised with a continuous default so that stock
applications get autofocus without sending AfTrigger.
Provenance and the adaptations made to each patch are recorded in the
patch commit messages.
Teaches the software ISP's AGC the IMX896 gain law
(gain = 16384/(16384-code), inverted from the vendor driver and
round-tripped against all 3009 vendor gain-table entries, max error
0.049%) and adds the sensor properties (1.0um unit cell, test-pattern
map). Same shape as the OV13B10 pair; both verified to apply and
compile on v0.7.2 after 0001-0004.
The extracted-CCM tuning yaml is deliberately NOT packaged yet - the
color matrices come from the vendor Chromatix blob and the licensing
call is open (see the camera journal).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Kernel config: CONFIG_VIDEO_OV13B10=m + CONFIG_VIDEO_DW9714=m — the camss
core, CAMCC and CCI were already enabled; the sensor and VCM drivers were
the only missing pieces. The camera carries themselves land via
milos-linux combined-stable (merge of the 'camera' topic branch,
nondescriptpointer's 5 patches, authorship preserved).
libcamera: new temp/ aport overriding pmOS 99990.7.2-r1 with -r2, adding
nondescriptpointer's OV13B10 sensor helper (gain = code/128, enables the
software-ISP AGC/AWB) and sensor properties patches on top of the pmOS
0.7.2 APKBUILD. Patch sha512s re-recorded from the actual files (his
APKBUILD diff carried stale hashes); all four patches verified to apply
against the v0.7.2 tarball.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
