IMX471 AIQB tuning forces AE to ISO 4096 in well-lit room

[mikeBoterf]

Environment: Sony IMX471 (SONY471A), Chicony CJFOE833 module, Lunar Lake IPU7X, ThinkPad X9-14 Gen 1, AIQB: IMX471_AAJH5-D_LNL.aiqb (IQStudio 24.40.2.0)

Problem: Auto-exposure converges at maximum gain in a normally lit room:

AE exp[0] ag 16.000000 dg 4.000000 Fn 2.400000 time 33334us iso 4096
AE Converged: YES, distance_convergence: -0.465956

Manual exposure at gain=18 (ISO 800) produces a properly exposed image. The AIQB tuning has incorrect sensor sensitivity calibration.

Impact: Produces extremely noisy video unusable for conferencing. Forces users to libcamera SoftISP to get working AE.

Additional findings:

• HAL NR (strength 20) has no visible effect at ISO 4096
• Increasing driver fll_def does not help — HAL tet limits cap at 33ms
• HAL CCM and color pipeline work correctly — only AE is broken

[jdapena]

Same root cause on a different sensor and platform — OmniVision ov08x40 on Panther Lake (IPU7 / ipu75xa), Dell XPS 16 (DA16260). I think this is the same AIQB sensor-sensitivity-calibration defect you describe, so adding a second data point here rather than opening a duplicate.

Symptom (identical)

In a normally-lit room the auto-exposure converges at maximum analog gain with the exposure capped, producing video that's too noisy to use. Same picture as your ag 16.0 … iso 4096.

Setup

• Sensor: ov08x40 (ACPI OVTI08F4), module KAFE799
• HAL: libcamhal-ipu75xa + gstreamer1.0-icamera (icamerasrc), build 0~git202604141000 (PPA oem-solutions-group/intel-ipu7 / Dell somerville)
• Tuning: OV08X40_KAFE799_PTL.aiqb, graph OV08X40_KAFE799.IPU75XA.bin
• Sensor config config/linux/ipu75xa/sensors/ov08x40-uf.json: single full-res MediaCtlConfig, supportedISysSizes [[3856,2176]]

Auto-AE operating point (read live from the ov08x40 v4l2-subdev while streaming 1080p)

exposure       value=2250   (max 4992)      ≈ ~30 ms, exposure is NOT maxed
analogue_gain  value=1984   (min=128=1x → ~15.5x, this is the MAX)
digital_gain   value≈1024   (≈1x)

So the AE leaves >half the exposure range unused and pins analog gain at its maximum instead. mean luma ≈ 128–135 (correct brightness, all via gain).

What I verified

• icamerasrc gain-range and exposure-time-range are ignored in auto AE — the operating point above is identical for every value I set.
• The AE does not extend exposure even at 15 fps (where the sensor reports max exposure 9992 lines) and even with exp-priority=shutter — it stays at ~2250 lines and maxes gain. Matches your "HAL temporal exposure time limit ~33 ms".
• Manual AE reaches the hardware fine: at the analog-gain floor the image is clean but dark (mean ~36); reaching correct brightness requires the high gain → so the sensor is healthy, the AE/sensitivity calibration is wrong.

Strong evidence it's the full-res sensitivity calibration

The ov08x40 driver also exposes a 2×2-binned 1928×1088 mode. I added a second MediaCtlConfig (id:1, 1928×1088) to ov08x40-uf.json. With binning active (media-ctl shows SGRBG10_1X10/1928x1088), the analog gain immediately dropped to the floor (value=256) at the same target brightness — i.e. with ~4× more light per pixel the miscalibrated AE no longer over-gains. (PSYS then fails because the graph has no binned input node: Failed to update tuning / PipeManager bindExternalPorts failed to bind input port / configure psys dag failed, so 0 frames — but the gain drop is the point.)

Cross-check with Windows (same laptop, no NPU/Studio Effects)

Clean image at analog gain = minimum. OBS offers up to 2560×1440, but 1440p is visibly blurrier than 1080p → Windows reads the binned 1928×1088 mode and upscales. So the Windows cleanliness comes from binning (low gain), not from a noise-reduction step (NR has no effect at max ISO here either).

Reproduction

Find the ov08x40 sub-device (note the /dev/v4l-subdevN whose entity is ov08x40 X-0036):

media-ctl -p -d /dev/media0 | grep -nE 'ov08x40|v4l-subdev'

Stream 1080p with default auto-AE in one terminal, and read the live sensor regs + active ISYS mode in another:

# terminal 1
sudo gst-launch-1.0 -e icamerasrc device-name=ov08x40-uf \
  ! video/x-raw,format=NV12,width=1920,height=1080,framerate=30/1 ! fakesink
# terminal 2 (while it streams)
v4l2-ctl -d /dev/v4l-subdevN --list-ctrls | grep -E 'exposure|analogue_gain|digital_gain'
media-ctl -p -d /dev/media0 | grep -m1 -oE 'SGRBG10_1X10/[0-9]+x[0-9]+'
# -> exposure≈2250 (of 4992), analogue_gain=1984 (MAX), digital≈1024 ; ISYS = 3856x2176

gain-range / exposure-time-range are ignored, and 15 fps does not extend exposure — the regs above are unchanged for all of:

sudo gst-launch-1.0 -e icamerasrc device-name=ov08x40-uf gain-range=0,6 exposure-time-range=10000,33000 \
  ! video/x-raw,width=1920,height=1080,framerate=30/1 ! fakesink
sudo gst-launch-1.0 -e icamerasrc device-name=ov08x40-uf exp-priority=shutter \
  ! video/x-raw,width=1920,height=1080,framerate=15/1 ! fakesink

Manual AE does reach the hardware (sensor is healthy; min gain = clean but dark):

sudo gst-launch-1.0 -e icamerasrc device-name=ov08x40-uf ae-mode=manual exposure-time=30000 gain=0 \
  ! video/x-raw,width=1920,height=1080 ! videoconvert ! jpegenc ! multifilesink location=/tmp/m_%03d.jpg
# regs: analogue_gain near floor (256); captured frames are dark (mean luma ~36)

Binned-mode check: add a second MediaCtlConfig to config/linux/ipu75xa/sensors/ov08x40-uf.json (the driver rounds any sub-1928 size to the 2×2-binned 1928×1088 mode), routing 1080p to it:

"id": 1, "configMode": "AUTO", "output": [1928,1088], "format": "V4L2_PIX_FMT_SGRBG10",
   formats/link/videonode as id:0 but width=1928 height=1088
supportedStreamConfig: 1920x1080 -> "mcId": 1
supportedISysSizes: [[1928,1088],[3856,2176]]

then re-run the 1080p stream:

sudo gst-launch-1.0 -e icamerasrc device-name=ov08x40-uf ! video/x-raw,width=1920,height=1080 ! fakesink
# media-ctl -> SGRBG10_1X10/1928x1088 ; analogue_gain drops to 256 (floor) at the same brightness
# but PSYS then fails (graph has no binned input node):
#   CamHAL[ERR] PipeLine: Failed to update tuning 4
#   CamHAL[ERR] PipeManager: @bindExternalPorts, failed to bind input port
#   CamHAL[ERR] ProcessingUnit: @configure configure psys dag failed

Request

The OV08X40_KAFE799_PTL.aiqb AE/sensitivity calibration makes the AE converge at max analog gain at full resolution — the same defect as this issue. Any of these would fix it, lowest-effort first:

1. Correct the sensor-sensitivity calibration in the .aiqb (would fix full-res directly).
2. Lift/expose the ~33 ms HAL exposure-time cap so long-exposure / low-fps can trade gain for exposure.
3. Ship the 1928×1088 binned graph + config + tuning for ipu75xa/ov08x40 (sensor + ISYS + AE already support it; only the PSYS graph node is missing).

Happy to provide full logs, the patched json, or run any test that helps.

[jdapena]

Follow-up on my ov08x40 / Panther Lake (ipu75xa) data point: I tracked the root cause to the aiqb being an uncalibrated placeholder, and confirmed the calibrated tuning exists and works on the Linux HAL.

The shipped Linux aiqb is a placeholder

OV08X40_KAFE799_PTL.aiqb as committed in intel/ipu7-camera-hal (config/linux/ipu75xa/) and shipped by the libcamhal-ipu75xa-common package:

• IQStudio 24.15.4.0, 456907 bytes, md5 e02b5ea2efc12583c6764116793a82a3
• Internal changelog has only 3 entries: "Initial tuning with copy from Mersey", "Add 4C1D 3A tuning", "Add NVM in LSC and AWB" → it was never module-calibrated.
• (The Dell OEM Linux repo dell.archive.canonical.com/.../somerville ships a byte-identical file, so this isn't specific to the public PPA.)

The calibrated tuning exists — in the Windows driver for the same module

From the Windows IPU7 camera driver (ov08x40.inf DriverStore), same module, same filename OV08X40_KAFE799_PTL.aiqb:

• IQStudio 25.35.4.0, 448078 bytes, md5 04368acec567e21847589988a2ddd137
• Build tag [PTL][POR][OV08x40][1C/DOL] 25ww45_... (POR), with ~1 year of real calibration in the changelog: Re-calculate CMC, less_noise gain 8-62, several update TNR, update LSC, enable flicker detection v2, new ISP/pipes, HDR, etc.

The calibrated aiqb works on the Linux HAL and fixes the over-gain

I dropped the Windows 25.35.4.0 aiqb in place of the placeholder on Linux (libcamhal-ipu75xa + libia-*-ipu75xa from the Feb-2026 build). Result:

• ABI-compatible — the HAL/icamerasrc initialises and streams fine, no crash.
• Auto-AE analog gain, same scene/lighting, full-res 1080p:
  • placeholder 24.15.4.0 → analogue_gain pinned at 1984 (max, ~15.5×)
  • calibrated 25.35.4.0 → analogue_gain ~1408 (~11×), plus visibly better TNR/CCM.
• So the placeholder's incorrect sensor-sensitivity calibration was indeed driving the AE to max gain (this issue's symptom). The calibrated tuning measurably reduces it. (Remaining gap to Windows is that Windows also runs a 2×2-binned readout — see my earlier comment — which the ipu75xa graph doesn't expose, so on Linux we're still full-res at ~11× in dim light.)

Request

Please ship the calibrated OV08X40_KAFE799_PTL.aiqb (IQStudio 25.35.4.0, the one already used by the Windows driver) in the Linux ipu75xa package / intel/ipu7-camera-hal config, instead of the 24.15.4.0 "copy from Mersey" placeholder. It's demonstrably ABI-compatible with the current libia and directly reduces the AE over-gain this issue is about.

Separately: the binned 1928×1088 mode (the rest of the gap)

The remaining difference vs Windows is that Windows runs a 2×2-binned readout (≈4× more light/pixel → AE keeps analog gain at the floor → clean), while the Linux ipu75xa config only exposes full-res 3856×2176 readout. I confirmed the sensor/ISYS support binning: adding a 1928×1088 MediaCtlConfig to ov08x40-uf.json makes the sensor switch to SGRBG10_1X10/1928x1088 and the analog gain immediately drops to the floor — but PSYS then fails (bindExternalPorts failed to bind input port / configure psys dag failed) because the Linux graph OV08X40_KAFE799.IPU75XA.bin has no binned node (the value 1928 appears 0× in it; only 3856/2176).

The binned graph demonstrably exists for this exact module: the Windows driver ships graph_settings_OV08X40_KAFE799_PTL.bin (5.8 MB vs the Linux 351 KB) in which 1928 appears 8656×. But it can't be reused on Linux — the Linux gcss "dispersed" graph reader can't parse the Windows-format graph (GraphConfig: failed to init graph reader; header format version differs). Unlike the aiqb, the graph is OS-stack-specific.

So please also generate/ship the 1928×1088 binned graph in the Linux gcss format (+ the matching MediaCtlConfig / supportedISysSizes, maybe I can PR). That, together with the calibrated aiqb above, could bring the Linux camera in line with Windows.

[jdapena]

Ideally what is needed is actually to obtain the .bin (compatible with Linux) and .aiqb for all the cameras in this repository, up to date, including binning, and adapting the json files accordingly. Not sure if anything else would be needed.