What orbis-reframe knows about the cameras' recorded files, kept up to date with its development. On GitHub.
Insta360 X6 file format
Desk research (2026-10-15), then two contributors' clips, read with our
code on their machines: a test set (2026-10-18: 24 clips, 8K / 6K / 4K,
Standard, I-Log and Dolby Vision, firmware v1.2.36) and another camera
(2026-11-11: same firmware, 8K 30 Standard with deliberate turns, Dolby
Vision, I-Log, 8K 50). The ✅ below are from those clips; the maintainer
hasn't examined an X6 file yet. Public specs, Insta360's manuals and
Gyroflow's telemetry-parser for the rest. Written as differences
from the X5 (insta360-x5.md, "X5 §n" below). Markers as
in README.md.
Files and streams
📖 One
.insvper take holding both lenses, the X5's naming (X5 §1).📖 Two 1/1.1" square sensors (a third more area than the X5's). H.265 only, up to 180 Mb/s.
📖 360 modes:
mode frame fps per lens (⚠️) 8K 7680×3840 50 / 48 / 30 / 25 / 24 3840² 6K (new) 6016×3008 60 / 50 / 48 / 30 / 25 / 24 3008² 4K 3840×1920 100 / 60 / 50 / 48 / 30 / 25 / 24 1920² ✅ Clips at all three lens sizes open and stitch. The app names a clip's mode by its lens size: 3008² is 6K (the X6 row of
orbiscore::InsvModel).✅ Vulkan decode + stitch on an AMD Radeon 8060S (Strix Halo iGPU, RADV), 8K 30: ~84 fps decode, ~35 fps for the full 7680-wide stitch, ~60 fps for the direct view.
❓ 8K 50 in real time (~1.5 Gpx/s across both lenses, ~1.7× the X5's 8K 30, X5 §3.3). A contributor's
playback_bench(60 frames) on an NVIDIA RTX 5070 Ti gave decode ~30.9 fps, fused view ~16.9 fps, direct view ~18.4 fps. That is far below the AMD figure above at the same frame size, and below an X5 8K 30 clip on an RTX 5070 (decode ~125 fps, fused view ~94 fps, 2026-11-11), so something else likely limits that run; not found yet. A stand-in on the RTX 5070 (2026-11-11; an X5 8K 30 clip's lenses re-encoded to HEVC Main 10, 3840², 50 fps, P frames only, ~76 and ~140 Mb/s per lens, Releaseplayback_bench --frames 60): decode 188 / 160 fps, fused view 118 / 106, direct view 117 / 104 -- 5x real time. Frame size, 10-bit and 50 fps don't explain the contributor's figure, nor a Debug build (the same) or CPU decode (the bench reports nothing off the all-GPU path, and that run did). Still ❓ until an X6 8K 50 file runs here: its encoder settings (B frames, tiles, bitrate peaks) may differ from the stand-in's.
Modes and sample clips
All the X6's modes 📖 (Insta360's shooting specs, Sources). We have no X6 file; the contributor's set is the 24 clips of 2026-10-18, the contributor's camera the 2026-11-11 clips (both fw v1.2.36). The status words are the X5's (X5 §3.4), and enough needs a clip we can run ourselves.
| Mode | Spec | Files | Clips | Status |
|---|---|---|---|---|
| 360 Video 8K | 50 / 48 / 30 / 25 / 24 | one .insv + .lrv |
contributor's set; contributor's camera: 8K 30 (turns), 8K 50 | 8K 50 wanted (real-time decode, see above), and any 8K clip of our own |
| 360 Video 6K | 60 / 50 / 48 / 30 / 25 / 24 | as above | contributor's set | one clip of our own wanted |
| 360 Video 4K | 100 / 60 / 50 / 48 / 30 / 25 / 24 | as above | contributor's set | one clip of our own wanted |
| I-Log 10-bit | as above | contributor's set; contributor's camera: 8K 30 moving | confirmed on contributors' clips (opens, stitches, limited-range BT.709, log profile); one of our own wanted, moving | |
| Dolby Vision (HLG) | as above | contributor's set; contributor's camera: 8K 30 | tagging and export with the shipped HLG LUT confirmed on contributors' clips; one of our own wanted: the HLG LUT's roll-off | |
| Lens guards (Standard, Premium) | as above | none known (the contributors' clips are bare: no field 145) | one clip with each guard, and the same scene bare, wanted | |
| IMU axes | contributor's camera: 8K 30, deliberate turns | "z,-x,-y" confirmed on a contributor's clip (see Trailer); one of our own wanted to check it with our own fit |
||
| PureVideo | 8K / 4K, 30 / 25 / 24; 6K at 5760×2880 | ❓ | none | wanted: 2880² lenses, which the X6's modes don't list |
| InstaFrame 2.0 | 360 8K 30 / 25 / 24 + flat 4K | .insv + .mp4 ("360 Video Backup" on), .mp4 only (off) |
none | wanted |
| Timelapse | 16K 15360×7680, 8K, 30 / 25 | ❓ | none | wanted: 16K is 7680² per lens |
| TimeShift | ❓ | none | wanted | |
| Bullet Time | 8K 7680×1920 100, 6K 6016×1504 120, 4K 3840×960 240 | ❓ | none | wanted: not a 2:1 frame |
| Loop Recording, Road Mode | ❓ | none | wanted, low | |
| Single-lens Video, PureVideo | 5K / 4K / 2.7K, 4:3 to 9:16, up to 120 | .mp4 |
none | wanted |
| Single-lens FreeFrame | 6016² 30 / 25, 3040×3840 60…25 | PRO_VID_….mp4 |
none | wanted |
| 360 Photo (HDR, AEB, Interval, Starlapse) | 120 MP 15520×7760, 30 MP | .insp, .dng, .jpg |
none | wanted, low |
| Audio: Auto, Engine Sound Enhancement, Auto Wind Noise Reduction, Voice Enhancement, Stereo, 360 Audio | 48 kHz AAC | none | one clip of each wanted |
Colour: 10-bit, log and HDR
- 📖 Colour modes Standard, Dolby Vision and I-Log 10-bit.
- ✅ Every clip is H.265 Main 10 (
yuv420p10le), in all three colour modes. Standard and I-Log clips are taggedbt709; Dolby Vision clips are taggedbt2020nc+arib-std-b67(HLG), limited range, with no Dolby Vision RPU side data. - ✅ Standard and I-Log streams are limited range on the contributor's
camera's clips (FFmpeg
tv/yuv420p10lewithout full-range), unlike the X5's full range (X5 §3.3). The app decodes each clip with the matrix and range its stream is tagged with; a stream without a range tag is taken as limited, the HEVC default. - ✅ The mode is only in the trailer, record 0x101 field 22
gamma_mode:standard,I_LogorDolby_Vision(X5 §3.3 for the X5's values). The app reads it as the clip's colour profile; onlyI_Logis recognised as a log profile so far. Media information shows Dolby Vision as Dolby Vision (HLG) when the stream is HLG. - The app's GPU stitch samples the decoder's pictures through a ycbcr
conversion. Since 2026-10-18 it takes 10-bit HEVC (Main
10) too: FFmpeg's Vulkan decoder gives P010 images
(
VK_FORMAT_G10X6_B10X6R10X6_2PLANE_420_UNORM_3PACK16), the stitcher's conversion is made for that format (by the stream's bit depth), and the viewport's shared lens planes are 16-bit (R16 / R16G16). Checked on a stand-in -- the X5 I-Log clip re-encoded to Main 10 with its trailer kept: export (both GPU paths, and the CPU fallback) and playback match the 8-bit original at 46-48 dB PSNR, in real time -- and ✅ on the contributor's X6 clips. Since 2026-10-19 the conversion takes the stream's matrix and range (BT.709 / BT.601 / BT.2020, full or limited) on the GPU, in the viewport and on the CPU path alike. Checked on stand-ins -- an X5 clip re-encoded to BT.2020 limited range, 8- and 10-bit: 45-47 dB PSNR against the original, where the old fixed conversion gave 29 dB -- and ✅ export of a real Dolby Vision clip through the shipped HLG → Rec.709 LUT (a contributor's, 2026-11-11,reframe_export). - An I-Log clip stitches correctly but looks flat until Insta360's LUT is
applied (below). Dolby Vision clips are HLG / BT.2020 as tagged and need
an HLG -> Rec.709 LUT; Insta360 publishes none, so the app ships one
(
tools/make_hlg_lut, from the BT.2100 / BT.2408 formulas) and offers it for any clip whose stream is tagged HLG (arib-std-b67). - The X5 has I-Log too, 8-bit (X5 §3.3).
- 📖 Insta360's 10-bit I-Log LUTs (
X6_I-Log_LUTs.zip, version 2.0, made 2026-05-20): 3D.cubefiles, each at 33³ and 65³, from I-Log to Rec.709 / BT.1886 (X6_I-Log_to_Rec709_BT1886_s65_v2.cube), to Rec.2020 PQ and to Rec.2020 HLG (1000 nits, 203 nits diffuse white). The app reads all of them (Colour ▸ LUT); the Rec.709 one is the one for its SDR output. The 10-bit pictures are graded in float before the output, so they shouldn't band; a 10-bit export (Project settings ▸ Bit depth) keeps 10 bits through to the file.
Trailer
- ✅ The X5's framing (magic, footer, record directory). Record 0x101 has
serial, model (
Insta360 X6) and firmware; the IMU (0x300, ~1 kHz) and exposure (0x400) records parse with the X5's layouts. - ✅ The extended MEI calibration string is there, the X5's 56-token
layout: e.g. xi 2.455, fx / fy 7232, cx / cy ~3850 / 3880 (the
contributor's set) and fx / fy ~7220, cx / cy ~3865 / 3889 (the
contributor's camera) -- per unit, from the factory -- roll ~90,
combined frame 15488×7744, type 193. The sanity ranges in
Calibration.cpphold. It is not in the short maker-notes tags, soinsv_trailer_dump'sParameters:line is empty; the string is on itsExtended:line. - ✅ Field 27
window_crop_info: src 7744², dst 7680² (X5: 5376² -> 5312², X5 §5.1). The dst width is the window fx / fy are given in, so the stitch reads it from the file (InsvMakerNotes::crop_dst_w->LensCalibExt::focalRefW); with the X5's 5312 the lens circles come out too small and the sphere has gaps between them. A file without the field takes its model's value fromorbiscore::InsvModel, with a log line. - ✅ No field 145 lens-profile table in any contributor's clip (all bare lenses): no lens-guard or dive-case profiles, so the stitch takes every lens as bare at the default 200° field of view. Stitches at all three sizes cover the sphere without gaps, so 200° looks right for bare lenses.
- 📖 The X6 has lens guards as accessories, in a Standard and a Premium version, like the X5 (Insta360 store; retailers, Sources). A retailer says the camera detects fitted guards and adjusts its stitching.
- ❓ What a guarded X6 clip records: whether the trailer then has a field
145 table in the X5's numbering and layout (X5 §5.4:
ProtectorS,ProtectorA, ... with a θ→height polynomial and field of view each), the profiles under another field or name, or only a flag saying which guard was fitted. Until a clip with each guard shows it, the app stitches every X6 clip as bare lenses and greys out Lens guards; a guarded clip then stitches with a slightly wrong lens model. - ✅ IMU axis map
"z,-x,-y"(the X5's, X5 §6.3; the X6 row oforbiscore::InsvModel::imuAxes), measured 2026-11-11 by a contributor on an 8K 30 Standard clip (~27 s) with a still start, then deliberate turns about each axis, fw v1.2.36; to check on a clip of our own. Method:tools/residual_stab --diagnose(trim 4–26 s) and the same tool's residual pan on 4–12 s againstz,x,yand-x,y,-z.- Horizon lock vs gravity: mean tilt ~0.4° with
"z,-x,-y". - Stick-band correction / raw rotation ratio 1.02 at scale 1.
- Residual forward pan lowest with
"z,-x,-y"on the compared stretch. - Video↔︎IMU offset from the file's timing ~+380 ms (exposure lock-in
~+394 ms); gyro ~1 kHz, rad/s. telemetry-parser has no X6-specific
code (unknown models get
Xyz).
- Horizon lock vs gravity: mean tilt ~0.4° with
Sources
- Insta360 X6 manual, shooting specs: https://onlinemanual.insta360.com/x6/en-us/specs/shooting-specs
- Insta360 X6 manual, file formats: https://onlinemanual.insta360.com/x6/en-us/operating-tutorials/storage/fileformat
- Insta360 manual, I-Log mode (X5 and X6): https://onlinemanual.insta360.com/x6/en-us/operating-tutorials/capture-preview/features/ilog
- Insta360 X6 specs page: https://www.insta360.com/specs/x6
- Insta360 store, X6 Standard Lens Guards: https://store.insta360.com/ca/product/x6-standard-lens-guards
- Clifton Cameras, Insta360 X6 Lens Guards (Standard and Premium): https://www.cliftoncameras.co.uk/insta360-x6-lens-guards/kit-variation/premium
- George's, Insta360 X6 Standard Lens Guards (the camera detects them): https://www.georges.com.au/products/insta360-x6-standard-lens-guards
- Insta360 I-Log LUT downloads: https://www.insta360.com/download/i-log
- Gyroflow telemetry-parser,
src/insta360/mod.rs: https://github.com/AdrianEddy/telemetry-parser