Getting the Arducam IMX708 / Raspberry Pi Camera Module 3 working on a Jetson Orin Nano running JetPack 6.2 — with the exact, hardware-tested device-tree and CSI parameters that actually work, and a clear list of the things that look right but don't.
-76B900){kind=icon}
Why this repo exists: the stock RidgeRun /
jetson-io/nvarguscamerasrcpath that works on JetPack 6.0 fails on JetPack 6.2 (capture timeouts, missing ISP tuning, ignored overlays). This toolkit captures the specific configuration that gets a clean 4608×2592 @ 14fps capture on JP6.2, so you don't have to rediscover it — and it fixes the washed-out color most people hit, in software, since the IMX708 has no NVIDIA ISP tuning.
- The 60-second version
- Supported configurations
- What works vs. what does NOT work on JetPack 6.2
- Critical CSI parameters
- Hardware prerequisites
- Quick start
- Repository structure
- Documentation
- Capturing images
- Image quality & color
- Troubleshooting & FAQ
- Specifications
- Licensing
- Credits
- Contributing
On JetPack 6.2 the IMX708 works, but only with a hand-corrected device tree and raw V4L2 capture:
- Build the
nv_imx708kernel module fromjp62-imx708-rpi-v3/driver/. - Use the corrected overlay
docs/overlays/imx708-nvidia-csi.dts— not the driver's bundleddts/overlay, which ships RidgeRun's original CSI values and times out on JP6.2 (see Critical CSI parameters). - Pre-merge that overlay into the base DTB with
fdtoverlayand point theFDTline inextlinux.confat the result. TheOVERLAYSdirective is ignored by the UEFI boot flow on JP6.2. - Capture raw Bayer with
v4l2-ctland debayer in Python/OpenCV.nvarguscamerasrcneeds ISP tuning files that don't exist for the IMX708.
Full walkthrough: docs/installation.md.
| Camera | Platform | JetPack | Status |
|---|---|---|---|
| Arducam IMX708 12MP (RPi Cam Module 3) | Orin Nano Super | 6.2 (L4T R36.4.3) | ✅ Working |
| Arducam IMX708 12MP | Orin Nano | 6.2 (L4T R36.4.3) | ✅ Working |
| Arducam IMX708 12MP | Orin Nano | 6.0, 5.1.1 | ↩ Use RidgeRun's patches |
| Arducam IMX708 12MP | Jetson Nano | 4.6.4 | ↩ Use RidgeRun's patches |
The genuine Raspberry Pi Camera Module 3 and the Arducam IMX708 module use the same Sony IMX708 sensor and behave identically here.
Arducam now ships an official IMX708 driver for the Orin Nano, and RidgeRun's driver exists too — both get the sensor probing, but users frequently report poor image quality / color on JetPack 6.2, because the default ISP path has no IMX708 tuning. This repo's value-add is the corrected JP6.2 CSI parameters (so capture doesn't time out), CAM1-aware validation/diagnostic tooling, and a software color pipeline that tackles the quality problem directly (see Image quality & color) — plus the documented learnings behind all of it. If your sensor is already detected via Arducam's driver, you can still borrow this repo's raw capture + Python processing for better color.
| Component | Configuration |
|---|---|
| Driver | RidgeRun nv_imx708.ko with the sony,imx708 compatible string |
| Device tree | The corrected NVIDIA-CSI overlay, pre-merged with fdtoverlay |
| Boot | FDT line in extlinux.conf pointing at the merged DTB |
| Capture | v4l2-ctl --stream-mmap for raw Bayer (RG10) |
| Processing | Python + OpenCV with percentile normalization + gray-world white balance |
| Approach | Why it fails on JP6.2 |
|---|---|
The driver's bundled dts/…-cam1.dts overlay |
Ships RidgeRun's original CSI params (channel@1, discontinuous_clk="yes", lane_polarity="6") → uncorr_err: request timed out |
jetson-io "Camera IMX477-C" mode |
Capture timeouts, no frames |
nvarguscamerasrc |
Requires per-sensor ISP tuning files that don't exist for the IMX708 |
GStreamer v4l2src negotiation |
not-negotiated errors on the raw Bayer caps |
OVERLAYS directive in extlinux.conf |
UEFI boot ignores runtime overlays — you must pre-merge the DTB |
Simple bit-shift (>>2) normalization |
Washed-out / overexposed images; use percentile stretch |
The breakthrough was using NVIDIA's CSI parameters (from their IMX219/IMX477 overlays) with the IMX708 sensor timings, instead of RidgeRun's defaults:
| Parameter | RidgeRun default (FAILS) | NVIDIA value (WORKS) |
|---|---|---|
discontinuous_clk |
"yes" |
"no" |
lane_polarity |
"6" |
"0" |
| NVCSI channel | channel@1 |
channel@0 |
These corrected values are already applied in
docs/overlays/imx708-nvidia-csi.dts.
The camera still connects to the CAM1 physical port (I²C bus 9,
cam_i2cmux/i2c@1, sensor at 0x1a) — only the NVCSI channel node uses
channel@0.
| Component | Recommended | Notes |
|---|---|---|
| Board | Jetson Orin Nano Super (8GB) | The board this repo was tested on |
| Camera | Arducam 12MP IMX708 Wide-Angle | The exact module this repo was tested with. Any IMX708 module (incl. a genuine RPi Camera Module 3) uses the same sensor. |
| Ribbon cable | 15-pin (camera) ↔ 22-pin (Jetson) CSI FFC | 15cm/30cm; contacts face the heatsink |
| Storage | NVMe SSD or 64GB+ microSD | NVMe recommended for capture throughput |
| OS | JetPack 6.2 (L4T R36.4.3) flashed and booted | Download |
Assumes JetPack 6.2 is flashed, the camera is on CAM1, and you have build tools
installed (sudo apt install build-essential device-tree-compiler nvidia-l4t-kernel-headers nvidia-l4t-kernel-oot-headers).
# 1. Build the kernel module
cd jp62-imx708-rpi-v3/driver
./build.sh # builds src/nv_imx708.ko
sudo ./build.sh install # installs the module + autoload config
# 2. Build the CORRECTED overlay (not the bundled dts/ one)
cd ../..
dtc -@ -I dts -O dtb -o imx708-nvidia-csi.dtbo docs/overlays/imx708-nvidia-csi.dts
# 3. Pre-merge it into the base DTB (OVERLAYS directive is ignored on JP6.2)
sudo fdtoverlay \
-i /boot/tegra234-p3768-0000+p3767-0005-nv.dtb \
-o /boot/tegra234-p3768-0000+p3767-0005-imx708-nvidia-csi.dtb \
imx708-nvidia-csi.dtbo
# 4. Point extlinux.conf at the merged DTB via an FDT line, then reboot.
# (See docs/installation.md for the exact extlinux stanza.)
sudo reboot
# 5. Verify
ls /dev/video0
v4l2-ctl -d /dev/video0 --info
dmesg | grep imx708
jp62-imx708-rpi-v3/driver/scripts/validate.shPrefer scripts?
scripts/apply_overlay.shandscripts/fix_extlinux.shautomate the pre-merge +FDTwiring.
nvidia-jetson-toolkit/
├── README.md # This file
├── CLAUDE.md # AI-agent context / notes
├── LICENSE # MIT (project-authored files)
├── docs/
│ ├── installation.md # Full IMX708 installation guide
│ ├── ssh-setup.md # Headless SSH access to the Jetson
│ └── overlays/
│ └── imx708-nvidia-csi.dts # ✅ The working (corrected) overlay
└── jp62-imx708-rpi-v3/
└── driver/ # RidgeRun-derived kernel driver (GPL-2.0)
├── COPYING # GPL-2.0 license text for this subtree
├── src/ # Kernel module source (nv_imx708.c, mode tables)
├── include/ # Driver headers
├── dts/ # Bundled overlays (RidgeRun defaults — see note)
├── scripts/ # validate / diagnose / apply-overlay / fix-extlinux
├── build.sh # Native build helper
├── Makefile # Module + overlay build
└── README.md # Driver-specific build/usage notes
Note on
driver/dts/: those overlays contain RidgeRun's original CSI parameters and are kept for reference / other JetPack versions. For JP6.2 usedocs/overlays/imx708-nvidia-csi.dtsinstead.
| Document | Description |
|---|---|
| Installation guide | Complete, hardware-tested setup with troubleshooting |
| SSH setup | Configure headless SSH access to the Jetson |
| Driver README | Driver build/usage and the JP6.2 compatibility notes |
nvarguscamerasrc is unavailable (no ISP tuning), so capture raw Bayer with
v4l2-ctl and process it in Python:
v4l2-ctl -d /dev/video0 \
--set-fmt-video=width=4608,height=2592,pixelformat=RG10 \
--stream-mmap --stream-count=1 \
--stream-to=/tmp/frame.rawimport numpy as np
import cv2
# RG10 = 10-bit Bayer in a 16-bit container
raw = np.fromfile("/tmp/frame.raw", dtype=np.uint16)
img = raw[:4608 * 2592].reshape((2592, 4608))
# Percentile stretch handles varying lighting far better than a bit-shift
p2, p98 = np.percentile(img, [2, 98])
img_norm = np.clip((img.astype(float) - p2) / (p98 - p2) * 255, 0, 255).astype(np.uint8)
color = cv2.cvtColor(img_norm, cv2.COLOR_BAYER_RG2BGR)
cv2.imwrite("/tmp/capture.jpg", color)The installation guide has
gray-world white balance, continuous capture, and an IMX708Camera OpenCV
wrapper class.
Two things nearly everyone hits with the IMX708 on Jetson: washed-out or
off-color images, and the question "where's the camera_overrides.isp tuning
file?"
This repo does not use the NVIDIA ISP, so there is no camera_overrides.isp
— by design. nvarguscamerasrc and the hardware ISP need per-sensor tuning
files that NVIDIA doesn't ship for the IMX708. Rather than fight that, this
toolkit captures raw Bayer over V4L2 and corrects the image in software:
- Percentile normalization (2nd–98th percentile stretch) for exposure —
far better than a naive
>>2bit-shift, which looks washed out. - Gray-world white balance to remove the color cast.
That software pipeline is the color-tuning result — it lives in the Python in
docs/installation.md, not
in an .isp file.
If you specifically want the ISP / nvarguscamerasrc path (hardware color,
auto-exposure/AWB), you need an IMX708 ISP tuning file, which neither NVIDIA nor
this repo provides. The community starting point is to reuse an IMX477
camera_overrides.isp (under /var/nvidia/nvcam/settings/) and hand-tune it —
expect approximate color, since it isn't matched to the IMX708. That's the route
some people take on top of Arducam's official driver.
Maintainer TODO: multiple users have asked for a tuned IMX708
camera_overrides.isp. This repo intentionally ships the software pipeline above rather than an ISP tuning file. If a hand-tuned.ispexists on the capture Jetson, add it here (e.g. under anisp/directory) with install notes; until then, the software path is the supported answer.
uncorr_err: request timed out / capture hangs.
You're almost certainly on RidgeRun's original CSI params. Confirm your overlay
uses discontinuous_clk="no", lane_polarity="0", and channel@0 — i.e. the
docs/overlays/imx708-nvidia-csi.dts overlay, pre-merged into the DTB. Then
re-check the ribbon cable seating and orientation.
No /dev/video0 after reboot.
Run scripts/validate.sh or
scripts/diagnose_full.sh.
Check, in order: module loaded (lsmod | grep imx708), camera on CAM1
(sudo i2cdetect -y -r 9 should show 1a, or UU once the driver is bound),
and that the merged DTB is
referenced by the FDT line in extlinux.conf.
I added OVERLAYS … to extlinux.conf and nothing changed.
Expected — the UEFI boot flow on JP6.2 ignores the OVERLAYS directive. You
must pre-merge the overlay into the base DTB with fdtoverlay and point the
FDT line at the merged file. See Quick start.
nvarguscamerasrc errors out or gives black frames.
Also expected. nvarguscamerasrc requires NVIDIA ISP tuning files that don't
exist for the IMX708. Use the raw V4L2 + Python path above.
Camera not detected on I²C.
Verify it's on CAM1 (not CAM0), the ribbon contacts face the heatsink, and
the cable is fully seated with the latch closed. Power-cycle (full shutdown, not
just reboot). On CAM1 the sensor appears at 0x1a on bus 9 (via the
cam_i2cmux), not bus 2.
The kernel module loads but the camera still isn't detected.
"Module loaded" (lsmod | grep nv_imx708) only means the driver is resident —
it does not mean the sensor was found. If sudo i2cdetect -y -r 9 shows nothing
at 0x1a and there's no /dev/video0, the sensor never ACKed. Work through, in
order: (1) camera on CAM1, not CAM0; (2) ribbon contacts toward the
heatsink, fully seated, latch closed; (3) full power-cycle, not just reboot;
(4) confirm the merged imx708 DTB is actually booted (the FDT line) and run
dmesg | grep -i imx708. A probe attempt that then times out points at CSI
params/cabling; no probe line at all means the device tree isn't being
applied.
What exact camera did this repo use? An Arducam 12MP IMX708 Wide-Angle module (product page), on CAM1 with a 15-pin ↔ 22-pin CSI ribbon. Any IMX708-based module (including the genuine Raspberry Pi Camera Module 3) uses the same sensor and should behave identically.
Is the poor color — or 14fps — the camera's ceiling? No. The poor color is just the ISP-less raw output; the software pipeline in Image quality & color recovers it well. The single 4608×2592 @ 14fps mode is a limit of this driver's mode table, not the sensor — there are simply no binned modes exposed here.
Module fails to load / vermagic mismatch.
Rebuild against your running kernel: uname -r must match
modinfo src/nv_imx708.ko | grep vermagic. Reinstall
nvidia-l4t-kernel-headers / nvidia-l4t-kernel-oot-headers for the current
L4T and rebuild.
Images are washed out, too dark, or have a color cast.
This is the #1 IMX708-on-Jetson complaint and it's expected from the raw sensor
output. Correct it in software — percentile normalization (not >>2) plus
gray-world white balance — and tune exposure/gain:
v4l2-ctl -d /dev/video0 -c exposure=20000 -c gain=64. See
Image quality & color for why there is no
camera_overrides.isp file.
Can I use a different resolution or higher frame rate? This driver exposes a single mode: 4608×2592 at up to 14fps. No binned modes.
Does autofocus work? This raw-V4L2 capture path does not expose the IMX708 voice-coil focuser, so focus is whatever the module powers on to. Focuser control would require additional driver work.
Which JetPack does this target? JetPack 6.2 (L4T R36.4.3) specifically. For 6.0 / 5.1.1 / Jetson Nano 4.6.4, use RidgeRun's original patches.
| Parameter | Value |
|---|---|
| Sensor | Sony IMX708 |
| Resolution | 4608 × 2592 (12MP) |
| Frame rate | 14fps (max at full resolution) |
| Pixel format | RG10 (10-bit Bayer, RGGB) |
| Interface | CSI-2, 2 lanes |
| I²C address | 0x1a |
| I²C bus | Bus 9 (CAM1, via cam_i2cmux) |
This is a mixed-license repository:
- Project-authored material (documentation, examples, and scripts outside
jp62-imx708-rpi-v3/driver/) is available under the MIT License. - The bundled kernel driver in
jp62-imx708-rpi-v3/driver/is derived from RidgeRun'snv_imx708driver and the Linux kernel and is licensed under GPL-2.0, per the license headers in each source file. The full GPL-2.0 text is injp62-imx708-rpi-v3/driver/COPYING. - NVIDIA, RidgeRun, and Raspberry Pi copyrights and notices remain with their respective owners; see the individual source-file headers for the controlling terms where they differ.
- RidgeRun — original
nv_imx708driver and excellent Jetson camera support (driver source). - NVIDIA — the Jetson platform and the CSI overlay structure this build borrows from.
- Arducam — the IMX708 camera module.
- The Jetson community on the NVIDIA Developer Forums for shared debugging knowledge.
Found a fix, a missing step, or a config that works on hardware this repo hasn't covered?
- Open an issue describing your board, JetPack version, and symptoms.
- Submit a pull request — hardware-verified corrections are especially welcome.
- Share your results on the NVIDIA Developer Forums.
Tested on: Jetson Orin Nano Super 8GB · JetPack 6.2 (L4T R36.4.3) · Arducam IMX708