What orbis-reframe knows about the cameras' recorded files, kept up to date with its development. On GitHub.
Insta360 ONE X2 file format
Examined on two clips (2026-10-20).
Written as differences from the X5 (insta360-x5.md,
"X5 §n" below). Markers as in README.md. The ONE X2 facts
come from the clips below, read with ffprobe, insv_trailer_dump and a
throw-away trailer / protobuf walker; field names are telemetry-parser's
(X5 §13). The app reads the 5.7K two-file take since 2026-10-21 (the
calibration from offset_v3, the lens-1 file found beside the _00_ one,
telemetry in ms); what's still open is in §11 and §12.
The older, two-file generation: one file per lens at 5.7K. The X3
works the same way and has its own document,
insta360-x3.md.
Reference clips (sample_footage/insta-one-x2/, gitignored), one camera,
firmware v1.0.92_build1, recorded 2025-08-16 with the camera on a stick:
| take | files | length | notes |
|---|---|---|---|
103807_…_001 |
VID_…_00_001.insv, VID_…_10_001.insv, LRV_…_11_001.insv |
309.4 s, 7736 frames | |
105108_…_003 |
VID_…_00_003.insv, VID_…_10_003.insv, LRV_…_11_003.insv |
188.2 s, 4704 frames |
Twelve more takes from the same day and camera (added 2026-10-21, the same
three files each): _002, _005, _006, _008–_012 walking around a
bike meet on the stick (16–215 s), _004, _007, _013, _014 riding
(1323, 731, 419, 230 s). All are the same mode: 5.7K (2880² per lens) at
25 fps, H.264, the same offset_v3 and firmware, and every _10_ file is
one frame shorter than its _00_ (§3.1).
And one take from 2025-08-04 on older firmware (v1.0.62_build2, same
calibration): VID_20250804_152938_00_003 + _10_ + LRV, 5.7K at 29.97
fps (30000/1001), 1740.9 s (29 min), 52176 frames, 10.96 GB and 10.91 GB
-- the longest clip here. Here the two files have the same frame
count.
Seven more takes from that day and firmware (added 2026-11-08, the same
three files each), all 5.7K 29.97 fps: 152306_…_001 (6.9 s),
152716_…_002 (29.1 s), and five 1.7 s takes _005-_009 and one 0.97 s
take _010 -- edge cases for very short clips. Their _10_ files are one
frame shorter, as on the 25 fps takes (871 / 872, 50 / 51, 28 / 29), except
_001 (206 / 206).
Beside them are five flat videos exported from Insta360's software from
the same day (1.MP4, 2.MP4, 3.MP4, CKDH4299.MP4, EXAM7155.MP4):
3840×2160 30 fps HEVC Main 10, HLG (arib-std-b67, BT.2020) with a
Dolby Vision configuration record, tagged make Insta360, model
ONER.FixFrame, muxed by Lavf60.3.100. Edited (reframed, a tiny planet,
a title), so they show Insta360's look and HDR export, not a reference
frame for the stitch.
1. Files
- ✅ One file per lens:
VID_…_00_NNN.insvandVID_…_10_NNN.insv, the first digit being the lens (X5 §1). 📖 The X6 manual: "X6, X5, X4 Air and X4 generate only one VID file containing data from both lenses, unlike the X3, which produced two". - 📖 Only 5.7K writes two files. Insta360's file-format table: on the
X3, ONE X2 and ONE X, 360 video at 5.7K gives two
VIDfiles and oneLRV; below 5.7K (4K, 3K) oneVIDfile and oneLRV. ❓ How that one file holds the two lenses (two streams like the X5, or one). No LRV at all in Bullet Time, TimeShift, Timelapse, or video at 4K 60 / 50, 3K 100 or above 100 fps. - ✅ The proxy is
LRV_…_11_NNN.insv: extension.insv, not.lrv, and lens digit1(the X5 writesLRV_…_01_NNN.lrv). It holds both lenses (§3.2). - ✅ Only the
_00_file and the LRV carry theinsttrailer. The_10_file is a plain MP4 that ends in padding: no calibration, IMU or model name of its own. So the_10_file can't be recognised as an Insta360 lens file from its content; only its name and its_00_sibling say what it is. - ✅ The
_00_file's metadata (field 26file_group_info) doesn't name the_10_file:{type 6, index 0, identify ""}(index 1 on the LRV). 📖 telemetry-parser describes a sibling hint naming the other file of a pair; these files have none. It also warns that a long recording split into two files is a group too, so a pair can't be matched by name alone. ❓ How a split ONE X2 recording is named (no clip long enough).
2. MP4 container ✅
ftyp (32 B) major_brand "avc1", compatible "avc1isom"
mdat the A/V samples of this one lens
moov two tracks; udta holds only an Ambarella "AMBA" box (107 B)
free padding up to the next 1 MiB boundary of the file
trailer _00_ and LRV only: the inst records, raw, to EOF
- No
instbox header and nowidebox: the trailer starts right at the 1 MiB boundary afterfreewith its first record's payload. It's found from the footer at EOF (X5 §4.1), as on the X5.ExtractRawTrailer()("everything aftermoov") hands the walker the padding as well, which is harmless because the walk starts at EOF. - The
_10_file ends in itsfreebox; its size is an exact multiple of 1 MiB (1 761 607 680 = 1680 MiB). - Metadata field 9 (
file_size) is the trailer's start offset (the size before the trailer, padding included), as on the X5. - ✅ The container's
creation_timeis camera-local time marked UTC (Z), and 9-10 s before the time in the file name (110155→2025-08-16T11:01:45Z, on all 15 takes). The trailer has no recording time (field 7 is absent, §6), so the app takes the file name's. - No thumbnail in
moov(the X5 has a ThumbnailTIFF there). TheAMBAbox's content is unknown ❓; it holds the frame count (7736 / 7735). - Track timescale 25000; handlers
Ambarella AVC/Ambarella AAC, encoder tagAmbarella AVC encoder(X5:INS.*).
3. Streams
3.1 VID_…_00_ / _10_ (one lens each) ✅
| index | what | details |
|---|---|---|
| 0 | the lens's fisheye video | H.264 Main, avc1, level 5.2, 2880×2880, 25 fps, ~45 Mb/s, no B-frames (has_b_frames=2 in the header, but only I and P frames), GOP 16 (0.64 s) |
| 1 | audio | AAC-LC, 48 kHz stereo, 128 kb/s |
- Pixel format:
yuvj420p, 8-bit 4:2:0 full range like the X5, but tagged BT.601 (smpte170mprimaries, transfer and matrix), where the X5 says BT.709. Read the matrix from the stream (it's tagged), don't assume the X5's. - The fisheye circle doesn't fill the frame: the calibration radius is ~1481 px in the 3040² sensor frame, so ~1403 stream px in the 2880² frame, with a black margin around it (X5: the circle touches the frame edge).
- The
_10_file is one frame shorter than_00_on every 25 fps take (7735 vs 7736, 4703 vs 4704, ...) and on most 30 fps ones; the same length on two 30 fps takes (152938_003,152306_001), so neither the frame rate nor the firmware decides it ❓; itscreation_timecan be a second later. ✅ Frame n of one file is frame n of the other;_10_only lacks the last frame (measured 2026-10-21): the seam match (EstimateSeamAlignment's strip score, no lens-1 adjustment) with_10_shifted by −2..+2 frames, at 10 points along each take and 12 more over_001's last minute, is best at shift 0 almost everywhere -- mean block NCC 0.63 at 0 against 0.56 / 0.56 at ±1 on_001, 0.59 against 0.56 / 0.52 on_003; the few points that pick another shift have flat scores (little texture or motion at the seam) and disagree with each other. On the 30 fps take, 10 points over its 29 min: shift 0 at every one (0.38 against 0.29 / 0.30). Both files start at pts 0. - Every file has the same audio: the AAC streams of
_00_,_10_and the LRV decode to identical samples (first 20 s compared, md5; on_014the whole 230 s at lag 0, the_10_/ LRV streams one AAC frame shorter). One of them is enough: the app plays and exports the_00_file's, and its decoder gives FFmpeg's samples exactly. Unlike the X5's dual-mono streams, the two channels differ (L − R is about 29 dB below the signal on_014), so the camera records real stereo. - 📖 Specs (Insta360): 360 video up to 5.7K 30 (5760×2880), also 4K 50 and 3K 100; H.264 or H.265, up to 100 Mb/s. Only 5.7K 25 H.264 is in our clips.
3.2 LRV_…_11_ (proxy) ✅
| index | what | details |
|---|---|---|
| 0 | video | H.264 Main, 736×368, 25 fps, ~14 Mb/s, GOP 15, no B-frames |
| 1 | audio | AAC-LC 48 kHz stereo, the same samples as the masters |
- Both lenses side by side, 368² each, not stitched. The left half is
the
_00_lens (on the X5 lens 0 is the right half). - One frame shorter than
_00_, like_10_. - It carries the full trailer: IMU, exposure and type 9 are
byte-identical to the
_00_file's; the metadata differs only in field 9 andfile_group_info.index(1); its thumbnails (§5) are a 736×368 frame.
3.3 Every mode, and our clips of it
All the ONE X2's modes 📖 (Insta360's shooting specs, Sources), with our clips and whether they're enough to build and check the reader on. The status words are the X5's (X5 §3.4): enough, one clip, wanted.
| Mode | Spec | Files | Our clips | Status |
|---|---|---|---|---|
| 360 Video 5.7K | 30 / 25 / 24 | _00_ + _10_ + LRV_…_11_ |
14 takes at 25 (walking and riding), one 29 min take and seven short ones (0.97-29 s) at 29.97, H.264, gamma_mode linear |
enough at 25 and 30 H.264; 24 wanted |
| 360 Video 4K | 50 / 30 | one VID file (+ LRV at 30 only) |
none | wanted: the one-file layout |
| 360 Video 3K | 100 | one VID file, no LRV |
none | wanted |
| HDR Video | 5.7K 25 / 24 | ❓ | none | wanted |
| H.265 encoding | none | wanted | ||
| Filters Vivid / LOG | none | LOG wanted: Insta360 publishes a LUT for it (ONE-X2-LUT-Final-V1.0.0.cube, 65³, the same file as the X3's); Vivid low |
||
| Timelapse | 5.7K 30 | no LRV | none | wanted |
| TimeShift | 5.7K 30 / 25 / 24 | no LRV | none | wanted |
| Bullet Time | 3K 100 | no LRV | none | wanted |
| Steady Cam (Ultra Wide Stabilization, 150°) | 1440p 50 / 30, 1080p 120 / 50 / 30; HDR, Timelapse, TimeShift at 1440p | .mp4 |
none | wanted: a flat clip, may open as one already |
| Split recording | ❓ (§12 question 3) | none split: _004 (1323 s, 7.5 GB per lens file) and 152938_…_003 (1741 s, 11 GB per lens file) are one file per lens each |
wanted if longer ones split | |
| 360 Photo (Burst, Interval, Night Shot, HDR) | 18 MP | .insp, .dng (RAW) |
none | wanted, low |
| InstaPano (270°), 150° wide-angle photo | 4320×1440; 1440p / 1080p | ❓ | none | wanted, low |
| Audio: Wind Reduction, Stereo, 360 Direction Focus | ❓ which ours used | one clip of each wanted | ||
| Lens guards | ❓ whether ours had them | one clip with, one without wanted | ||
| IMU axes | the two takes (stick swing) | a clip with deliberate turns about each axis wanted |
4. The inst trailer — framing ✅
- Footer as on the X5 (X5 §4.1): the same magic, version 3, the trailer length.
- No type-0 Offsets table. The sequential walk from the footer (X5 §4.2) reaches every record; the first record in the file (at the trailer start) is type 10.
- Records,
_00_001(309 s, 7736 frames), walked from EOF:
| hdr id | type | name | size | contents |
|---|---|---|---|---|
0x0101 |
1 | Metadata (protobuf) | 1 843 | §6 |
0x0200 |
2 | Thumbnail | 2 343 322 | one H.264 keyframe, 2880×2880: the _00_ lens's first frame (Annex B, an 18-byte NAL of type 26 before the SPS). LRV: 736×368, both lenses |
0x0300 |
3 | Gyro (IMU) | 8 725 080 | 56-byte samples, 500 Hz (§8) |
0x0400 |
4 | Exposure | 124 144 | 16-byte samples, one per frame (§9) |
0x0500 |
5 | ThumbnailExt | 2 000 058 | the same, for the _10_ lens: its first frame, 2880×2880 H.264. LRV: identical to its type 2 |
0x0900 |
9 | AAAData? | 94 608 | 48-byte records, ~6.4 Hz (§10) |
0x0a00 |
10 | Anchors | 35 | as on the X5: 7 types, all count 0 |
- Not present: 0 (Offsets), 7 (GPS; no GPS source on these clips), and the X5's 11, 22, 28, 29.
- The
_00_file's trailer is the only place that has the_10_lens's thumbnail.
5. Thumbnails
Types 2 and 5 are each a single H.264 IDR frame that FFmpeg decodes as a raw
h264 stream ✅: the _00_ and _10_ lens's first frame at full size (X5:
one stitched 1280×640 NV12 frame). The app's preview for a ONE X2 clip
(since 2026-10-21) decodes both, scaled to 720², and stitches them on the
CPU with the clip's calibration into a 1024×512 frame of the X5's form:
~0.27 s per clip on the sample takes, once (the media card caches it).
6. Type 1 — Metadata (protobuf) ✅
Values from _00_001; _00_003 differs only where noted.
| # | name (telemetry-parser) | ONE X2 value | notes |
|---|---|---|---|
| 1 | serial_number |
IXSE… (14 characters) |
|
| 2 | camera_type |
Insta360 ONE X2 |
|
| 3 | fw_version |
v1.0.92_build1 |
|
| 5 / 17 | offset / original_offset |
16-token short calibration | §7 |
| 9 | file_size |
trailer start offset | |
| 10 | total_time |
309 (188) | seconds |
| 14 | gyro |
56 bytes | the IMU sample at the first frame: its timestamp is field 24, then the 6 doubles of that type-3 sample |
| 18 | trigger_source |
1 = CameraButton | |
| 19 / 44 | dimension / resolution_size |
{2880, 2880} | per-lens stream size |
| 20 | frame_rate |
25 | |
| 22 | gamma_mode |
linear |
X5: standard / I_Log. ❓ which colour profile writes linear |
| 24 | first_frame_timestamp |
4407 (4378) | milliseconds on the device clock (X5: µs) (§9) |
| 25 | rolling_shutter_time |
36.94 | readout, ms: nearly the whole 40 ms frame (X5: 11.8–21.2) |
| 26 | file_group_info |
{type 6, index 0, identify ""} | type 6 = VideoHdr in telemetry-parser's enum ⚠️; index 1 on the LRV |
| 27 | window_crop_info |
{src 3040×3040, dst 2880×2880} | sensor 3040² per lens; dst = the stream size, so the focal scale dimension/dst is 1 |
| 28 | gyro_timestamp |
−12.0 | ms; negative (X5: +1.6) |
| 29 | is_has_gyro_timestamp |
true | |
| 31 | gyro_calib |
6 × f64 + u64 | (−0.0034, −0.0090, +0.0090, 0.034, 0.038, −0.0053), then Unix time 1700149542 (2023-11-16; the X5 has 0 there) |
| 36 | first_gps_timestamp |
= field 24 | not a Unix time without GPS |
| 41 | is_selfie |
0 | |
| 43 | is_dewarp |
1 | |
| 45 | battery_type |
1 | |
| 46 | cam_posture |
0 | |
| 47 / 48 / 49 | fov_type / distance / fov |
3 = Ultrawide / 1.0 / 0 | default reframe view |
| 50 | gyro_filter_type |
2 | |
| 51 | gyro_type |
1 = InsdevImuType40609 |
the same IMU part as the X5 ⚠️ (enum value only) |
| 52 | media_data_rotate_angel |
1 | |
| 53 / 55 | offset_v2 / original_offset_v2 |
34 tokens | §7 |
| 54 / 56 | offset_v3 / original_offset_v3 |
40-token MEI | §7 |
| 57 | focus_sensor |
3 = All | |
| 58 | expect_output_type |
0 | |
| 60 | photo_rot |
9 floats | an exact rotation matrix (orthonormal, det 1), different per clip (167° and 170° from identity). ❓ Probably the camera's orientation at the start |
| 61 | audio_mode |
2 = Stereo |
- Absent compared with the X5: 7 (
creation_time), 11 (GPS), 42, 62 (is_raw_gyro, so false), 64 (pts_type), 65 (gyro_cfg_info), 93, 98 (first / last frame times, frame count), 111 / 112 (the 56-token MEI), 120 / 121, 145 (lens-accessory profiles). - Without field 145 there are no guard profiles to pick from. ❓ How Studio handles a guarded ONE X2.
7. Calibration ✅
The ONE X2 carries the 40-token offset_v3 (X5 §5.3) but not the
56-token string (fields 111 / 112) the app's stitch requires. (The desk
research guessed it had no offset_v3; it has.) All strings are stated in a
6080 × 3040 side-by-side frame (2 × 3040²), lens type 41.
offset_v3 (field 54), per lens xi fx fy cx cy yaw pitch roll tx ty tz k1 k2 k3 p1 p2 w h lensType (X5 §5.3), trailing flag 199680 (0x30C00):
| xi | fx fy | cx cy | yaw pitch roll | tx ty tz | k1 k2 k3 | p1 p2 | |
|---|---|---|---|---|---|---|---|
| lens 0 | 1.72859 | 2322.61 2322.41 | 1528.43 1521.45 | 0.799 −0.157 178.977 | 0 0 0 | 0.2279 −0.1489 −0.9684 | −0.00185 0.00116 |
| lens 1 | 1.72859 | 2318.48 2318.72 | 4554.65 1542.81 | −0.848 −0.368 1.421 | 0.00017 −0.00195 −0.0211 | 0.2436 −0.2279 −0.8443 | 0.00129 −0.00042 |
- Lens 1 sits ~21 mm behind lens 0 (X5: ~32 mm).
- The roll is ~180° on lens 0 and ~0° on lens 1 (X5: ~90° on both). In the X5 this angle is also the lens ↔︎ IMU rotation (X5 §6.3), so the ONE X2's IMU sits differently. The lens pictures themselves lie on their side (the horizon runs top to bottom in both files).
- ⚠️ With the X5's MEI reading (radial terms in r², r⁴, r⁶), the short string's radius (1481 px) falls at θ ≈ 100°: about a 200° lens, like the X5.
- What's missing for the app's MEI model: no
k4and no thin-prisms1..s4(the 56-token string has both); everything else maps one to one. Scaled to the stream,cx/cy · 2880/3040(lens 1:cx − 3040first) andfx · 2880/2880. offset(field 5), 16 tokens, per lensradius cx cy yaw pitch roll(X5 §5.2):2 1480.960 1532.890 1516.290 0.712 -0.131 179.140 1480.280 4551.640 1546.940 -0.762 -0.108 1.275 6080 3040 3113.offset_v2(field 53), 34 tokens, per lensradius cx cy yaw pitch roll tx ty tz+ 4 polynomial coefficients (1.0 −0.152 0.271 −0.112) +6080 3040 41, flag 134144 (0x20C00).- The
original_copies (17, 55, 56) equal the working ones. - ✅ The stitch takes it as the X5's string (checked 2026-10-21 on
_001,orbis_stitch --lens1):cx / cyscaled by 2880 / 3040 (half the 6080 frame),fx / fyby 2880 / field 27's dst 2880, and the same roll handling as the X5 (the stored roll less 90°). The stitched frame is continuous. Not yet compared with Insta360 Studio's stitch of the same frame. - The seam fine-tune (measured 2026-10-23,
_001,_004,_007,_013, one camera): far content is 0.24–0.46° off at the seam with the file's calibration alone, more than on the X5. Lens 1 is tilted (pitch −0.6 to −1.2°; the yaw varies between takes, −0.9 to +0.5°, traded against the centre, which looks alike at the seam), its centre +0.07 to +0.13 % of the width across and −0.03 to −0.12 % down, both lenses' radius −0.05 to −0.06 %, roll −0.08 to −0.09°. Fitted per clip, the offset on other frames of the take is 0.07–0.10°. - With no lens-profile table (field 145) the stitch takes the bare lens at its default 200° field of view.
- 📖 telemetry-parser's rig code mentions a ONE X2 lens type 42 (a lens guard whose optical table differs from its mask table). Our clips say 41, presumably the bare lens ⚠️.
8. Type 3 — IMU ✅
- 56-byte samples:
u64timestamp, then 6 ×f64— accel x y z in g (mean magnitude 1.01 g over the clip), then gyro x y z, ⚠️ rad/s (the doubles layout, X5 §6.1; field 62 is absent).ReadInsvTelemetry'slen % 56guess picks the layout correctly. - Timestamps in milliseconds (X5: µs), exactly 2 ms apart: 500 Hz
(X5 ~1006 Hz, irregular). The app reads them at the model's tick
(
InsvModel::telemetryTickUs, 1000) and hands them on in µs, like the X5's;insv_trailer_dump's summary does the same (its--imu/--exposurecolumns stay raw). - Pre-roll: the IMU starts 1.60–1.63 s before the first frame on the
2025-08-16 takes, 1.32 s on the 30 fps take (older firmware). The X5's
sync formula with the ms units gives a video ↔︎ IMU offset of +1619.7 ms
on
_001(residual_stab), within that range. - 📖 Axes: telemetry-parser has
xZyfor the ONE X2, but only in its branch for files withoutoffset_v3. These files have it, so it would use its defaultXyzafter rotating by lens 0'syaw pitch roll. - ✅ Ours:
"z,x,y"(the ONE X2 row oforbiscore::InsvModel::imuAxes; measured 2026-10-21): sphere X ← raw z, Y ← raw x, Z ← raw y -- the X5'sz,-x,-yturned 180° about sphere X. Method as X5 §6.3: four horizon views (yaw 0 / 90 / 180 / 270°, 90° FOV) rendered unstabilized byreframe_export, the camera rotation per frame interval from tracked features, put into the sphere frame and least-squares fitted against the gyro integrated over the same interval. Here the rotation is a rotation-only (Kabsch) fit of the tracked rays, not an essential matrix: the takes have slow and stationary stretches where the essential matrix degenerates. The chain was checked on the X5's193105first (essential matrix; it findsz,-x,-y, 5.6° from the free fit).- On 30 s of each take (
_001100–130 s,_00340–70 s):z,x,yis 0.4° (_003) and 3.6° (_001) from the free fit, the runner-up 90° away; every view on its own picks it (1–7°). The rotation left per frame: 0.11 / 0.14° against 0.74 / 0.55° raw. - The gyro is in rad/s: the free fit's scale is 0.95–1.00.
- The accel shares the gyro's axes (as on the X5): carrying the
low-passed accel along the integrated gyro, the identity ranks first on
both takes (weak evidence: the rotations are mostly about gravity). And
in the sphere frame it agrees with the picture: at 105–125 s of
_001it puts the camera's forward 28–34° below level, the unstabilized forward view's vanishing point ~30°. - Riding (added takes):
_01460–90 s givesz,x,y1.4° from the free fit, 0.13°/frame left of 0.82 raw._013120–150 s ranks it first too, but weakly (0.57 of 0.67°/frame): a ride with near parallax doesn't suit the rotation-only fit.
- On 30 s of each take (
9. Type 4 — Exposure and sync ✅
16-byte samples:
u64timestamp in ms,f64exposure time in s (0.62–1.31 ms on these sunny clips).Two samples at the IMU's start, a ~1.3 s gap (~1.07 s at 30 fps), then one every exactly 40 ms to the end (at 29.97 fps: 33 and 34 ms in the nominal 63:37 mix of 1001/30 ms rounded to ms): no drift at all at ms resolution, where the X5's measured 15–40 ppm would show (7–12 ms over 309 s). So these are probably nominal times, and the clock rate can't be fitted from them.
Field 24 is not an exposure timestamp (X5: it equals one exactly). It lies 5 ms before the nearest one on both clips (4407 vs 4412, 4378 vs 4383).
✅ The frame's time is field 24 + field 28 (−12 ms), as in the X5's formula (measured 2026-10-21): fitting the image rotation of the four views (§8) against the gyro over a lag sweep puts frame 0 at 4394 ms (
_001) and 4364.5 ms (_003) on the IMU clock, 13 ms before field 24; the app's offset (+1619.7 / +1588.7 ms after the first IMU sample) puts it at 4392.7 / 4363.7 ms. The app's sync fine-tune (residual_stab --auto-sync,stitch::MeasureClipSync) finds +1.1 / +0.8 ms on the same ranges, and +2.6 ms on the 30 fps take (600–630 s, a calm stretch: 13 cells), so the formula holds on the older firmware and at 30 fps too. The readout time (field 25, 36.9 ms) is used as read; it wasn't measured on its own.Counting from the first exposure at or after field 24: 7750 samples for 7736 frames (4714 for 4704): 14 / 10 after the last frame.
With field 24 (ms), field 28 (−12 ms) and field 25 (36.9 ms) the X5's sync formula (X5 §7.1) carries over once the units are ms; per telemetry-parser, files without
is_raw_gyroare already in seconds after the field-24 shift. Checked against the picture: see above.⚠️ The lenses' colour (measured 2026-10-31 from the picture, seam cells both lenses see alike, unclipped): unlike the X5's and the A1's, the difference wanders within a clip (
_001: lens 0 against lens 1, red +0 to +23 %), and even without flare the ratio along the seam is widely spread (on_00950-80 % of the cells more than 10 % off the frame's median, X5 / A1 10-30 %). Over a clip lens 0 is the brighter, red +11-12 %, blue +1-12 % (_001,_009). Not known why: separate auto exposure or white balance per lens (one exposure record serves both), or lens shading. The app matches the clip's average and follows the drift with a small, slow per-frame correction.
10. Type 9 ❓
- 48-byte records, a
u32ms timestamp first, about every 160 ms (155–165; 1971 records on 309 s), 30 all-zero records among them. - The X5 layout (X5 §4.4) doesn't fit the next fields (
ev_target~1e-26), exceptluma_struct(word 4):>> 19gives 64–65, so ISO 100 by the X5 formula.
11. What the app reads, and what it still needs
Read (2026-10-21):
The take's two files as one clip's lenses: the
_00_file (it has the trailer) is the clip, and its lens 1 comes from the_10_file beside it, taken only when that is a one-stream video of the same size and rate, at most a frame shorter, without a trailer of its own. The clip's length is the shorter file's. The proxy is found asLRV_…_11_….insv. Any of the three files names the clip (camera::ClipFileOf: the name says which_00_file, which must claim it), so importing, dropping or adding the_10_file or the LRV, or the whole folder, gives the clip once. The project saves the_10_file with the clip's media item ("files"), and the clip is missing when either file is.The stitch from the 40-token
offset_v3(field 54; §7).Telemetry in ms, turned into µs (§8).
The preview for the media card, stitched from types 2 and 5 (§5).
The matrix from the stream (BT.601 here), as for every camera.
The recording time from the file name (2026-10-25; no field 7, §2).
The IMU axis map
z,x,y(§8) and the X5's sync formula with ms units (§9), both measured on these clips; the two files paired frame n with frame n (§3.1).
Still needed:
- The stitch checked against Insta360 Studio's, and the axis map confirmed on a clip with deliberate turns about each axis.
12. Open questions
Answered 2026-10-21: the two files are frame-locked (§3.1); the frame's
time is field 24 + field 28 (§9); the axis map is z,x,y and the gyro is
in rad/s (§8).
- Which instant field 24 marks relative to the exposure record (5 ms apart) -- the sync doesn't need it (§9).
- What
gamma_modelinearandfile_group_info.type6 mean for the mode set on the camera; the other modes and the lens guards are untested (§3.3). - How a long recording is split into files: not by 4 GB or by 22 min
(
_004, §1), nor at 29 min / 11 GB (152938_…_003, §1). - Type 9's layout; the
AMBAbox; field 60. - Whether a frame is taken half its exposure before its time, as on the
X5 and the A1 (X5 §7.1). The per-clip sync measurement on 45 windows of
the 14 sample clips scatters by ±0.9 ms whatever the exposure (mostly
0.2-1.4 ms, its timestamps whole ms), so it doesn't tell; the app keeps
one lag,
field 24 + field 28 − 2.3 ms, and the per-clip fine-tune.
Sources
- Our clips (above), 2026-10-20:
ffprobe,insv_trailer_dump, a trailer and protobuf walk following X5 §4–§5. - Insta360 X6 manual, file formats (one file vs. two): https://onlinemanual.insta360.com/x6/en-us/operating-tutorials/storage/fileformat; the same table in the X5 manual has the X3 / ONE X2 / ONE X rows (two files at 5.7K only, when no LRV is written): https://onlinemanual.insta360.com/x5/en-us/operating-tutorials/storage/fileformat
- Insta360 X-series shooting specs, codecs and colour profiles per model: https://onlinemanual.insta360.com/onex2/en-us/specs/shooting-specs
- Prevu3D, both
.insvfiles needed per recording: https://doc.prevu3d.com/en/realityplatform/dataset-preparation-and-upload/input-types/photogrammetry-data/360-video/insta360-capture-setup/ - Insta360 LUT downloads (the ONE X2's LOG LUT): https://www.insta360.com/download/i-log
- Clifton Cameras, ONE X2 spec sheet (modes)
- Gyroflow telemetry-parser,
src/insta360/mod.rs(IMU orientation tables and when each applies, timestamp units, lens profile fromoffset_v3),extra_info.rs(field names and enums) andrig.rs(one-lens-per-file layout, sibling hint, lens type 42): https://github.com/AdrianEddy/telemetry-parser