What orbis-reframe knows about the cameras' recorded files, kept up to date with its development. On GitHub.
Antigravity A1 file format
Desk research (2026-10-15); measured on clips 2026-10-29: three
contributed 8K H.265 flights, then a second set from the same drone: one
H.264 clip in each of nine video modes (8K, 5.2K, 4K 100) and two photos;
a third set (2026-11-01, measured 2026-11-02): a take-off and a
landing, 5.2K 60 in H.265, 4K Slow Motion, a timelapse, a 14 MP photo and
AEB DNGs, and Antigravity Studio's 360 exports of one clip (Modes and
sample clips).
What's marked ✅ below is checked on them; the modes we have no clips of
are still desk research, from reviews and Gyroflow's telemetry-parser. The
app supports the A1 experimentally from 2026-10-29: 8K clips open,
stitch, stabilize (leveled by the drone's own attitude) and export; 5.2K
and 4K 100 clips open and stitch (below). The A1 is a 360 drone from Antigravity, a
brand incubated by Insta360, and writes the same .insv family; written as
differences from the X5 (insta360-x5.md, "X5 §n" below).
Markers as in README.md.
Streams
📖 Two 1/1.28" sensors. Reviews say 180° lenses, but ✅ each lens images ~200°, like the X5's: by the clip's own calibration the image circle's edge lies at ~100° off-axis all round (99-101° on lens 1; 101-105° on lens 0, where the frame clips the circle), measured on a frame of
_011and_012by where the picture goes dark. The stitch's 200° default (X5 §5.4, used when the file has no lens profiles) fits. The A1's frame is cropped closer to the circle than the X5's: the circle nearly fills the 3840² stream.📖 360 modes, H.264 or H.265, up to 170 Mb/s:
mode frame fps per lens (✅) 8K 7680×3840 30 / 25 / 24 3840² 5.2K (new) 5248×2624 60 / 50 / 30 / 25 / 24 2624² 4K 3840×1920 100 / 30 / 25 / 24 1920² ✅ on our clips: 8K at 30 (29.97, 30000/1001), 25 and 24; 5.2K at 60 (59.94, 60000/1001), 50, 30 (29.97), 25 and 24; 4K at 100 (exactly 100, stored at the capture rate); 4K Slow Motion, captured at 100 and stored at 29.97 (Slow motion and timelapse). Every mode writes the same calibration (below), so the per-lens sizes are the one 5248² sensor window scaled. The camera's target bitrate is field 135, in units of 2²⁰ b/s: 170 for 8K (both codecs), 105 for 5.2K, 140 for 4K 100; the two streams together measure 177 / 109 / 147 Mb/s (decimal).
✅ One
.insvholds both lenses as two square streams, HEVC Main or H.264 High, 8-bit, full range (yuvj420p), BT.709 on all our clips (handlerINS.AVCfor H.264), lens 0 (looking down) first, lens 1 (looking up) second. No audio track; amov_texttrack holds only a sensor frame counter per frame (F:13776, ...).✅ File names:
VID_<date>_<time>_<n>.insv, with no lens token (the X5's isVID_…_00_…), and the proxy beside itPRX_<date>_<time>_<n>.prx: the same name withPRXand.prx, both fisheyes side by side at 2560×1280 (the X5's.lrv, X5 §3.2), with a trailer of its own (records 1 and 2). ✅ Lens 0 (.insvstream 0) is the left half, where the X5's is the right: on_011at 10 s, each lens's frame scaled to 128² against each half of the proxy's, mean difference 4.7 levels against 93-99 for the other half (the same check on an X5 clip finds its lens 0 on the right, 5.6 against 47). ✅ The MP4'screation_timeis the recording start in UTC to the second: on our first three clips it is the name's time, which is local, four hours ahead (five on the 2026-11-01 set, the day the contributor's clocks went back).✅ The file holds its own name: in the MP4 header,
moov/udta, in thefreebox after theAMBAbox, as the path on the card,C:\DCIM\CAM_001\VID_<date>_<time>_<n>.insv(the proxy:PRX_….prx; the.insvalso holds its proxy's name). So a renamed clip can be given its name back, take number and all; the second set came renamed by mode and got its names back so, matching the time rebuilt fromcreation_timeand field 7 on every file. The app takes a renamed clip's recording time from it and finds the proxy by it. The photos hold no name.✅ A proxy carries its clip's field 24 (the first frame's time on the drone's clock) and serial, on all nine pairs of the second set: what ties a renamed proxy to its clip; the app takes one found by the held name or the clip's stem only when both match.
✅ Untouched files can be told apart. Field 9 of record 1 is the byte size of the file before the trailer, exactly, on every
.insvand.prxwe have (12 clips, 9 proxies): a re-muxed or trimmed file no longer matches it.✅ The drone in the picture: on our clips, only a light bar with a dark edge at the very rim of lens 0 (the downward one), 94-97° off its axis, in the same place on every clip -- the drone's body or an arm seen past the horizon. It lies at the outer end of the seam's blend band, where lens 0 hardly counts (the edge fade), so it doesn't show in the stitched picture (checked at the spot, depth-aware seam off and on). Nothing of the drone near the seam elsewhere, nor in lens 1. ✅ On the ground (take-off
_024, landing_025, seen in the proxies): two white landing legs reach into lens 0's picture from its rim, with the pad below them; ❓ how much of them the stitched picture shows (not checked).
Trailer
- ✅ The same 78-byte footer as the X5 (X5 §4.1): magic
8db42d694ccc418790edff439fe026bf, trailer version 3. The footer's header names the Offsets table (type 0, 570 bytes: 57 entries). - ✅ The Offsets table (X5 §4.2) is the only index. Each record's payload
starts a 256 KiB slot (the entries' offsets are 0x40000 apart, a few
records packed between them), and no six-byte header follows a
record: the bytes after it are zeros or the next record's data. So the
walker takes each record's type and size from its table entry, as
telemetry-parser does; reading them from the header after the record
finds nothing. Checked on all three clips and the
.prxproxy. - ✅ Records in the clips: types 1 (Metadata, protobuf), 2 (thumbnail), 3 (IMU), 4 (exposure), 7 (GPS), 9, 11, 14, 23, 32, 33, 37 (attitude), 38 and 56; the proxy holds 1 and 2. Of the ones the X5 doesn't have, 37 is read and 14 / 56 partly understood (below); 23, 32, 33 and 38 hold timestamped floats, meaning ❓ (a flight log?).
- ✅ Record 1 is the X5's protobuf (X5 §5.1); on our clips: model (field 2)
Antigravity A1, firmware10.0.397.3_5.5; field 27{5248, 5248, 5248, 5248}, so the focal reference is 5248; the 40-tokenoffset_v3MEI calibration (field 54, X5 §5.3; no 56-token string), in a 10496×5248 side-by-side frame,xi1.504, lens type 155, which the app reads; and a 34-tokenoffset_v2(X5 §5.3: radius, centre, yaw / pitch / roll, baseline, four polynomial coefficients; lens type 112) with the angles in another form, pitch near 0 (lens 0 yaw / pitch / roll 179.66 / −0.29 / 88.88, lens 1 179.90 / −0.43 / 89.69), not used. ✅ Both strings are identical on every clip of all three sets (17 clips, one drone), so the clips can't tell a misread calibration from a constant the vendor's software adds; fields 24, 25 (readout ~21.25 ms), 28 (3.7 ms), 62 (raw IMU) and 65 (±32 g, ±2000 °/s) as on the X5, and the IMU (~1 kHz) and exposure records decode with the X5's layout, in µs. Field 7 isYYYYMMDDhhmm, local time, without the X5's seconds, so the recording time comes from the name the file holds (Streams), then the one it has, which carry them. Field 195 is the capture rate and field 196 the stored (playback) rate, each a fraction (30000 / 1001, 100000 / 1000, ...); they differ only in slow motion (Slow motion and timelapse). Field 20 is the stored rate, rounded (30, 100); field 10 the recording's length in whole seconds, field 93 in ms (30730 on the 4K 100 clip, whose streams last 30710: the proxy's length) -- capture time, not video time, in slow motion; field 9 the file's size before the trailer (above); field 25 the sensor readout, ~21.2 ms at 8K, ~12.3 ms at 5.2K, 6.3 ms at 4K 100; field 132 the codec, 1 H.264, 2 H.265; field 135 the target bitrate (Streams). ⚠️ Field 153 looks like the recording mode: 0 on every 8K and 4K 100 clip, 3 on every 5.2K one (both codecs), 4 in 4K Slow Motion, 2 in the timelapse; 1 on the 55 MP and 14 MP photos, 5 on one 55 MP photo (_032, ❓ which photo mode). Field 26.1 is 0 on the normal clips, 17 in slow motion, 2 in the timelapse (the X5'sVideoTimelapse, X5 §3.5), 3 on photos. Fields 51 (1) and 64 (2) are set, meaning ❓. The serial is in field 1 as on the X5. - ⚠️ The calibration's extrinsics are yaw / pitch / roll ≈ (−177.5, −89.95, 90.4) and (−179.4, −89.9, 89.4): pitch −90° puts yaw and roll about the same axis, so the order they compose in matters (it doesn't on the X5). Pitch, then yaw, then roll (X5 §5.3) leaves the lens axes opposite to 0.2°, and the per-clip seam fit converges to 0.02° on all three clips with small corrections (lens 1 pitch +0.14 to +0.24°, roll +0.40 to +0.42°, centre −0.06 to −0.08 %, radius −0.17 to −0.20 %). Before the fit the lenses are 0.24-0.53° apart at the seam (X5: ~0.17°). The app's per-clip measurement gives the same on all three clips, over each whole clip (0.47-0.52° -> 0.02°). ✅ The same at every resolution: on the second set the lenses are 0.40-0.49° apart before the fit in all nine modes, and the fit finds the same correction at 8K 30, 5.2K 60 and 4K 100 (lens 1 pitch +0.23 to +0.27°, roll +0.43 to +0.54°, centre −0.07 to −0.09 %, radius −0.10 to −0.15 %), ending at 0.14-0.15°: these clips fly low, a few metres over a field and trees, so parallax is left.
- ✅ Studio's stitch puts lens 1 where the per-clip fit does, not where
the file alone does (measured 2026-11-03,
stitch_benchon_0214-8 s against Studio's unstabilized equirect, Modes and sample clips). The export lines up with our sphere atrefrot=0.361,-2.764,89.940(Studio puts the lens axes on the equator; from 1298 matched features at 6 s, median 0.036°, and the bench's own fit within 0.1°) andrefstart=0. Against it, our stitch from the file's calibration alone is 0.16° (median) / 0.57° (p95) off in the seam band, band PSNR 24.0 dB; with the per-clip seam correction (lens 1 rotation −0.002 / +0.221 / +0.430°, centre −0.07 %, radius −0.15 %) 0.14 / 0.23°, 25.5 dB. The difference is a tilt of lens 1 by ~0.48° against our reading ofoffset_v3, seen also 10-30° from lens 1's axis, out of reach of Studio's AI Stitching at the seam; lens 0 sits where the file puts it (radial within 0.01°), so Studio doesn't apply the fit's radius scale to it. The fit's lens 1 rotation alone matches Studio best, the same on 4-8, 30-34 and 60-64 s (band 0.10 / 0.34 / 0.30° median, p95 0.27 / 0.66 / 0.58°, against 0.16 / 0.37 / 0.34° and 0.57 / 0.77 / 0.77° from the file alone); adding its centre or radius terms makes it worse. So Studio's lens 1 holds one pose through the clip: the file's, turned by that rotation. No other order of the angles gives it (the closest,Rx Rz Ry, still leaves 0.27°). ❓ Where Studio's tilt comes from. - ✅ The app's new stitcher corrects the file's calibration by a fixed
amount for the model (
orbiscore::InsvModel::lensCorrection, as offsets from the file's values in the per-clip seam fit's terms, measured 2026-11-03): lens 1 turned (+0.006, +0.238, +0.468)° (yaw / pitch / roll on top of the file's angles), its centre moved (−0.0749, +0.0016) % of the lens picture's width, both lenses' radius −0.1286 %. Each is the median over all 17 clips of the per-clip seam fit (yaw −0.002 to +0.011; pitch 0.10 to 0.30, 0.59 on the landing_025; roll 0.41 to 0.54, 0.76 / 0.88 on_027/_024, −0.97 on_025; centre −0.071 to −0.084 % and −0.006 to +0.009 %; radius −0.08 to −0.21 %). The lenses are then 0.14° apart at the seam on average (17 clips, 8 frames each; 0.42° with the file's calibration alone, 0.30° with lens 1's turn only), and the seam is the sharpest of every correction tried (band sharpness over beside, on_010,_012,_015and three stretches of_021; tied on_024). A fit of all six numbers to every clip at once scores 0.10° but makes the seam less sharp than the median: the offset it minimises leaves out any shift the way parallax goes, and a larger radius change on both lenses turns the far content's error into that direction. Against Studio's_021export the median correction puts both lenses ~0.1° further out at 50-80° from their axes than Studio, which keeps the file's radius (lens 1's turn alone matches Studio closer: band 0.09 / 0.33 / 0.29° median against 0.13 / 0.35 / 0.33°); from the seam alone, which lens's radius is off can't be told -- a chart calibration of an A1 would. ⚠️ One unit measured: if the offsets are this drone's own calibration error rather than the model's, another A1 needs other values; a clip from a second A1 would tell. - ✅ The lenses are 71 mm apart (lens 1
tz−0.0711; X5: 32 mm), lens 0 in front along its own axis as on the X5: the depth-aware seam lines up houses, poles and wires along the horizon when the drone flies low (_011at 1 s, confidence 0.89), and with the sign flipped (orbis_stitch --baseline -0.0711) they double again. - ✅ A low sun on the seam (2026-10-30,
_021at 60-70.4 s, flying high toward a setting sun): the up lens sees the sun at its rim with its own glare and rays, the down lens a weaker, different flare. Matched as plain grey values, the two lenses agree badly at the true (far) distance there (NCC 0.25-0.45, ~0.95 elsewhere on the seam), and the glare's smooth gradient lines up with itself at any shift, so the depth-aware seam took 0.3-0.4 m and bent the town round the sun; at the 71 mm baseline 0.3 m is a 13° shift (X5: 6°). The seam's match now takes each lens's brightness plane out first (orbiscore::SeamPatchScore), and those bins stay far on all the frames checked. The table also stays within the X5's shift range (−0.9 to +5.9°), which at 71 mm ends at ~0.7 m, and a match has to be clearly better than any other shift to count (the uniqueness term,orbiscore::SeamDepthTuning::uniqueLo/uniqueHi): a frame solved from itself after a seek (_018at 22.36 s) took scattered matches at 0.3-0.5 m and pulled the far town into a loop. - ✅ The lenses' colour (measured 2026-10-31 from the picture, seam
cells both lenses see alike, unclipped, over each clip): the exposure
record (type 4) holds one shutter time per frame, not one per lens,
and there is no type 12: the lenses expose together, and the file
doesn't say how they differ. Measured, the difference is fixed within a
clip (±1.5-3 % from frame to frame outside flare; on
_010it doesn't follow the shutter, 0.84-1.30 ms, within ±2 %) but differs between clips: the down lens (lens 0) against the up one, blue −1.2 % on_010, −6 to −8 % on the second set (_021R −0.9 %, G −2.4 %, B −6.2 %). Across the overlap it changes by +0.3 to +0.5 % per degree towards lens 0: lens 0 falls off towards its rim more than lens 1, most in blue. With the sun at the up lens's rim (_02123-33 s, 57-70 s) that lens's veiling flare adds light (lens 1 ≈ 0.56-0.87 × lens 0 + 26-90 levels on the sun's side): such frames show as a wide spread of the ratio along the seam, and the app's per-clip measurement (orbiscore::FitLensColour) leaves them out. ⚠️ Not the same all round the seam within a frame: on_020at 3-5 s (flying low, the sun near the horizon) the lenses agree within 1-5 % on half the seam circle, while on the ~135° facing the sun the down lens is 15-45 % darker (most in blue); and away from the sun the per-clip value (G −2.4 %) is ~2 % off on stretches where the lenses agree. The stitch therefore also matches them per stretch of the seam, per frame (orbiscore::SeamBlend::local). ⚠️ Veiling light is added light, not a gain (measured 2026-10-31 on_015at 23.9 s, the sun just above the horizon at the up lens's side): on the seam stretch under the sun the up lens sees shadowed grass at G 107 against the down lens's 58, and the sky next to it a little brighter -- in linear light (levels^2.2) both about +0.10, a ratio of 2.6 on the grass and 1.7 on the sky. No gain matches both; taking the added light out does (orbiscore::LensVeilPerBin,SeamBlend::veil: up to 0.25 linear there). On the same frame the sky's blue is near clipping (245-250) in both lenses while red and green agree within 1-2 levels: with whole cells left out for one clipped colour, that stretch had no data and was given its neighbour's correction (a 14 % step). The lens colour is measured per colour since (orbiscore::kLensColourMethod2). ⚠️ The difference depends on the content (measured 2026-10-31 on_020at 2.17 s, turning): in the sky just above the trees the lenses agree within 1-2 levels, in the trees' shadows below lens 1 is 8-12 levels higher in blue -- a small added offset, under what counts as veiling light. Matched on all the overlap's rows, the sky split by 4-8 % in green on a line; the stitch matches each stretch on the rows where the lenses meet, and follows the picture within ~40 ms, as the scene slides along the seam when the drone turns. - ✅ How the lenses sit in the picture (
InsvModel::mountDeg, the stitch's output turn): not in the file -- the calibration places the lenses relative to each other, as on the X5, where lens 0 faces forward. Stitched the X5's way, the A1's lens axes lie on the horizon. Turned so lens 0 looks straight down and the picture's forward is the drone's nose: yaw −90°, pitch 90°, roll 0 (the X5's: 180, 0, 0). The nose found from the horizon on 2026-10-29: a drone flying fast tilts nose down, so the horizon in the stitched picture is highest toward the nose. A fitted sinusoid on six frames of clip_010at a steady 15.7 m/s (62-72 s, tilt 16-19°) puts the high point at −86° ± 6° from lens 0's side; taken as −90°, the body square to the lens frame. It puts the trailer's thumbnail (type 2) of_011at the picture's centre too;_012's thumbnail looks elsewhere, so the thumbnail is a view, not the nose. Measured from the scene (2026-11-05,_021): the direction of travel in the stitched picture -- the essential matrix of feature matches between two frames 10 apart, on the sphere's bearings -- against record 37's velocity taken into the picture through field 69, the 9.7° tilt and the IMU axes below. On the five pairs with little rotation (30.5, 50, 65-67 s; 12-14 m/s) they agree to −0.16° in azimuth (−0.87 to +0.81°) and +0.19° in elevation (−0.16 to +0.73°): the nose sits at the picture's yaw 0, and the elevation checks the tilt without the attitude-against-gyro fit it came from. Insta360 Studio's 360 export of the clip without FlowState isn't turned to the nose: it keeps its lens frame, the lens axes on the horizon, about 2.65° about lens 0's axis from ours (360NoStab.mp4of the contributor'sPanoExport). - ✅ IMU axes
-x,y,-z(InsvModel::imuAxes): sphere X ← −raw x, Y ← raw y, Z ← −raw z, in the sphere the A1's mount gives. Measured 2026-10-29 as on the X5 (X5 §6.3) and the ONE X2: four unstabilized horizon views (yaw 0 / 90 / 180 / 270°, 90° diagonal field of view) of_010, the camera rotation per frame interval from tracked features (rotation-only Kabsch fit), against the gyro integrated over the same interval. On 40-60 s (turns at 100-150 m): 1.3° from the free fit, gain 1.01 / 0.97 / 0.99, every view picks it (1.7-2.7°), and the rotation left per frame drops from 1.20° to 0.074°; on 10-40 s the same map, 2.7°. The runner-up is 90° away. Raw +y is also where the accelerometer puts "up" in a hover (0.92 g of it). - ✅ Thumbnail, record type 2, in a layout of its own: a 64-byte
header --
u32 0,u32 64(the header's size),u32the picture's size in bytes, then zeros but for width (u32at byte 24), height (28) and row stride (32) -- then a flat view, not an equirect, NV12: 256×192 in the clips, 768×768 in the.prxproxy, the stride equal to the width in both (the X5's: a 40-byte header and a 1280×640 equirect, X5 §4.3). The view looks ahead and down at the ground, a wide lens look; on_011the proxy's picture shows the same view as the clip's, cut square. Checked 2026-10-29 on all three clips and their proxies. Full range like the video: 2-14% of the clips' pixels are above 235, a sky in the video's lens 1 likewise (10%), and the pictures look right decoded so. The matrix is ⚠️ BT.709, the video's, not measured: a view fitted from the stitched first frame matched no better under BT.601 or BT.709. The app shows the picture on the A1's cards (Media tab, project library), cut to the card's shape. - ✅ GPS, record type 7, in the X5's layout (X5 §8): a fix every
1.6 s on all twelve clips (the X5 with the GPS remote: once a second),
one record per fix, from the first one the drone has after recording
starts. Field 98 (X5 §5.1), which on the X5 gives the first frame's Unix
time, is absent, and nothing else in the file ties the GPS clock to the
video: field 36 is the first fix's own Unix time (to the ms on all
three), the MP4's
creation_timeis in whole seconds, and no other record holds a Unix time. - ✅ The GPS lined up with the video by its speed (
AlignGpsBySpeed,Gps.h): the fixes' speed against record 37's horizontal speed, timed in video time by field 24, over every shift; the best to the ms. On_010/_011/_012the first fix is 1.129 / 0.911 / 0.526 s into the video, matched to 0.20 / 0.06 / 0.08 m/s RMS, 0.8-2.0 m/s a second either way. A check that doesn't use the fit's own clip: the three put the drone's clock (field 24) at the same offset from Unix time to within 31 ms; the nine clips of the second set, one per mode, 24 to 100 fps, 15-70 s at 0.1-15 m/s, to within 11 ms. A shift is taken only when it is clear: at least 5 fixes over the attitude track, their speeds spread by 0.5 m/s or more, matched to 1 m/s RMS or better, and every shift a second or more away fitting at least twice as badly (+0.1 m/s). A hover, or a clip too short for 5 fixes (8 s), stays without GPS in titles and gauges, and the log says why. - ✅ Sync: the X5's formula (field 24 + field 28, X5 §7.1) holds; the app's per-clip fine-tune finds +0.9 to +1.2 ms on all three clips, and the image fit above agrees within ~1 ms. Field 24 is one of the exposure record's timestamps, as on the X5 (all 13 clips checked).
- ✅ The centre row half the exposure before the frame's time, as on the X5 (X5 §7.1; measured 2026-11-05): the per-clip sync measurement on three 4 s windows of 16 clips (46 windows, every mode, exposure 0.4-5 ms, readout 6.3-21.3 ms) falls by 0.484 ms per ms of exposure, 0.10 ms RMS left (0.78 ms with one constant), at the X5's −0.5 ms; the readout time adds nothing.
- ✅ Attitude, record type 37 (the walker's id 0x2500): the flight
controller's orientation, 50 Hz, 36 bytes each:
u64µs on the IMU's clock (4-8 ms behind the gyro), a unit quaternion as x, y, z, w (4 × f32), body to world -- body x the nose, y right, z down; world north, east, down -- then the velocity north, east, down in m/s (3 × f32; its horizontal speed is the GPS speed). In fast flight the body pitches −25° to −28°. Measured 2026-10-29 on all three clips by matching the attitude's rotation per 20 ms step against the gyro integrated over the same step: body ← IMU is field 69's rotation, transposed -- (w, x, y, z) = (0.5, 0.5, −0.5, 0.5), an exact axis permutation -- then a 9.5-9.9° pitch up about the body's right axis, the camera tilted in the airframe (26° body pitch − 9.7° = the 17° the camera's horizon tilts). 0.09-0.14° per step is left. The tilt isn't in the file; the app takes 9.7° (InsvModel::attitudeTiltDeg, one drone measured). Insta360 Studio's FlowState export of_021(2026-11-05, 13 frames, each lined up with our stitch by feature matches) holds the attitude's up within 0.3° in steady flight and moves up to 0.65° off it, in pitch, in turns of 10-20°/s: no offset fixed in the airframe, so nothing that would point to a different tilt. - ✅ The accelerometer reads the drone's acceleration with gravity. In
steady flight it points up however the drone is tilted (thrust and drag
balance gravity:
_01062-64 s at 15.6 m/s, 0.2-1.3° from the attitude's up), but the drone speeds up, brakes and turns most of the time, which moves it 4-35° off up (2 s means,_010). The attitude's up plus record 37's velocity, differentiated, predicts it to ~1° (0.3-4° per 2 s), which checks the attitude and the tilt above. Leveled by the accelerometer, Horizon lock's up was up to 15° (pitch) and 9° (roll) off per 10 s on_010, 5-8° on_011, under 1° on_012; leveled by the attitude, 0.2° or less on all three (residual_stab --diagnose, block [6]). The app levels by the attitude (Horizon lock, Direction lock, Lean follow's auto-level). - ✅ The attitude's heading doesn't drift, and its north is its own.
Its yaw matches its own velocity's course in flight (median 0.0°, 10-90%
−1.8° to +2.5° over 5 m/s,
_010): the nose points where the drone flies. Against the GPS course (which matches the course from the GPS positions, median 0.1°) it is a steady 22.5-23.9° off on all three clips, minutes apart -- so record 37's north (and its velocity's) is the flight controller's own, not true north, and not magnetic north either (the declination there is ~13° the other way). For holding a direction only its steadiness matters. Against it the gyro alone, integrated, turns 2-5°/min (_0112.2,_0123.6,_0104.8), and the stabilizer's levelled track on_01017.7° over the 87 s (its bias fit can't see the part about up). The app holds the levelled track's heading to the attitude's north (ImuTrack::orientation):_010within ±0.5° at 11 s steps, 1.6° peak to peak with fast turns' transients (the attitude at 50 Hz, 4-8 ms late), so Direction lock doesn't wander. - ⚠️ Other orientation records (layouts from the bytes, meaning inferred, not used): type 14 has one 36-byte entry per video frame with a unit quaternion in it; type 56 holds the same orientation in another layout (its first quaternion is type 14's, sign flipped). ❓ Field 91 (a unit quaternion, none of the attitude, the IMU mount or their products).
Slow motion and timelapse
✅ One clip of each (2026-11-01, measured 2026-11-02). In both, the video's
time is not the drone's clock: the motion and exposure records run in
real time while the video plays slower or faster. The app turns video
time into the drone's clock by a capture scale, seconds of the clock
per second of video (orbiscore::CaptureScale): field 196 over field 195 in
slow motion (0.2997), a timelapse's frame interval times the stored rate
(60), from its frames' own times (type 6) where it has them, else from
field 30. Sync, stabilization (its smoothing windows stay in video time),
the GPS and the motion overlays all go through it (2026-11-04).
- 4K Slow Motion (
_028): two 1920² HEVC streams stored at 29.97 fps (2997/100), 2773 frames, 92.5 s of video. Captured at 100: field 195 is 100000 / 1000, field 196 29970 / 1000; the exposure record has one entry every 10.0 ms (2823), the IMU spans 28.2 s, field 93 is 27740 ms, and the readout (field 25, 6.33 ms) is 4K 100's. So the file is tagged at the playback rate and plays 3.34× slowed as it stands. Records as in a normal clip (types 1-4, 7, 9, 11, 14, 23, 32, 33, 37, 38, 56); a proxy as usual, also at 29.97. With the scale, the file's sync holds (the image's yaw left after Horizon lock is smallest at the file's offset, 0.053°/frame, against 0.114 unstabilized; ±20 ms both worse) and the frame clock fits to +0.1 ppm; the GPS, lined up by its speed, spans 3.0-88.4 s of the video (before: 0.9-26.5 s). ❓ How a trimmed slow-motion file's time map (record 0x8000) counts: taken as video time, untested. - Timelapse (
_029, 8K H.265): two 3840² streams at 29.97 fps, 150 frames (5.0 s) over ~299 s of flight: one frame every 2 s. As on the X5 (X5 §3.5), field 26.1 is 2, field 30 is 2000 and field 59 is true (the interval in ms); field 133 is 2000 too. Record type 6 holds oneu64timestamp per frame (150; the X5 writes one more), the first one field 24's, 2.002 s apart (2001.92-2002.08 ms: 60 frames of the sensor's 29.97), so the scale is 60.0, not the 59.94 field 30 gives. Unlike the X5's, the A1's timelapse keeps the IMU (type 3, ~1 kHz over the whole 299 s) and the attitude (37); the exposure record runs at 29.97 Hz over the whole time (8956 entries), GPS every 1.6 s. No proxy. With the scale, Horizon lock keeps it level (pitch left in the image 0.19°/frame against 0.70 unstabilized). Its GPS doesn't line up: the drone hovered (0-0.17 m/s), so there's no speed to match.
Photos
✅ Two 55 MP single shots (2026-10-29, the same drone), then two more and a 14 MP shot (2026-11-01):
14 MP (
_031): 5248×2624, the two fisheyes side by side at 2624², otherwise as below.DNG (AEB,
AEB_<date>_<time>_<n>.dng, three files, one per bracket, consecutive numbers): uncompressed 16-bit raw, RGGB, 10496×5248 -- both fisheyes side by side, as the.insp-- ~115 MB each, modelAntigravity A1. The three shutter times are 1/4722, 1/2258, 1/1129 s: 1 EV apart, ISO 100-105. ❓ Whether they hold the calibration (not checked).An
.inspis a JPEG, 10496×5248, the two fisheyes side by side (5248² each, lens 0 left as in the video), with a small Exif block: makerYingling Innovations Pte. Ltd., modelAntigravity A1, exposure, ISO, GPS position and altitude; no date and no Software tag (the X5 writes both).After the JPEG, the same
insttrailer as the video's: record 1 with the same calibration and fields (field 26.1 is 3, 0 on normal videos; field 91 is all zeros; no field 36, 93, 132 or 135), the IMU (record 3, 2 kB), and records 9, 11, 32, 38 and 56; no GPS record (the position is in the Exif) and no attitude (record 37). Field 9 here is 2329 bytes more than the JPEG's size on both photos, ❓ why.Unlike the video's, the photo's trailer also has the X5's six-byte header after each record (X5 §4.2), besides the Offsets table.
Thumbnail, record type 2, in the X5's 40-byte header layout --
u32 1,u32the record's size,u32 1,u32 0, width, height -- then a 256×192 NV12 flat view, the same kind of picture as the video's.
The app doesn't open 360 photos yet (roadmap, 360 photos); the trailer holds all the stitch needs.
Modes and sample clips
All the A1's modes 📖 (Antigravity's spec table as retailers and reviews
reprint it, and the spring 2026 update for timelapse; Sources). Our clips,
from one contributor and one drone (firmware 10.0.397.3_5.5): three 8K
H.265 flights (2026-10-28), and a set of nine H.264 clips, one per mode,
plus two photos (2026-10-29), flown low over a field. The second set came
renamed by mode but is otherwise straight from the drone (field 9,
Streams); the videos have their names back (Streams): _013 8K 25,
_014 8K 24, _015 5.2K 60, _016 5.2K 50, _017 5.2K 30, _018 5.2K
25, _019 5.2K 24, _020 4K 100, _021 8K 30.
The photos keep the contributor's names. A third set
(2026-11-01), with the drone's own names, over a park: _024 8K 30
H.264, take-off from a pad, then a hover with turns (43 s); _025 the
same, a low hover and landing (10 s); _026 its proxy only (67 s, on the
pad); _027 5.2K 60 H.265 (32 s); _028 4K Slow Motion; _029 a
timelapse; photos _030 / _032 55 MP, _031 14 MP; _033-_035
AEB DNGs. All our files come from the microSD card.
Studio references for _021 (8K 30, H.264): Antigravity Studio's
exports as 360 equirect, 7680×3840 H.265, ~200 Mb/s, at exactly 30
fps (2113 frames; the clip has 2111 at 29.97), so frames are matched by
time, not by number. Unstabilized: AI Stitching, FlowState off,
Chromatic Calibration off, no adjustments -- the reference for the
stitch's geometry, seam and colour. Stabilized: the same with FlowState
on, level. Each has a silent AAC track. The unstabilized one is still in
the drone's body frame and lines up with the clip from its first frame
(Trailer above has the rotation). Studio's earlier flat
3840×2160 exports of the same take (AI Stitching and Chromatic
Calibration on) show the raw body frame without FlowState (horizon on
end).
The status words are the X5's (X5 §3.4).
| Mode | Spec | Files | Our clips | Status |
|---|---|---|---|---|
| 360 Video 8K | 7680×3840, 30 / 25 / 24 | ✅ one .insv with both lenses (two 3840² streams), and a PRX_….prx proxy |
3 at 29.97 fps, H.265; 5 at 29.97 / 25 / 24, H.264 (with a take-off and a landing); standard colour | experimental: opens, stitches, stabilizes and exports; GPS lined up by its speed |
| 360 Video 5.2K | 5248×2624, 60 / 50 / 30 / 25 / 24 | ✅ as above, two 2624² streams | 5, one at each rate, H.264; 1 at 60, H.265 | experimental, as 8K: plays in real time (60 fps: 4.0 ms a frame on an RTX 5070), Horizon lock level within 0.2°, sync within 5 ms, exports (5.2K 60, _015) |
| 360 Video 4K | 3840×1920, 100 / 30 / 25 / 24 | ✅ as above, two 1920² streams; 100 fps stored at 100 | 1 at 100 fps, H.264 | as 5.2K: plays at 100 fps (3.3 ms a frame), level within 0.5°, exports (_020); wanted: 30 / 25 / 24 |
| 4K Slow Motion | 3840×1920, 30 / 25 / 24 | ✅ as 4K, stored at 29.97, captured at 100 (field 195; Slow motion and timelapse) | 1 at 29.97, H.265 | experimental: plays as stored (3.34× slowed, as the drone means it); sync, stabilization and GPS on the capture time (by the tools); Media information shows Slow motion |
| H.264 encoding | up to 170 Mb/s, as H.265 | ✅ as above, H.264 High; field 132 = 1 | 11 | as its mode in H.265 |
| Timelapse | added by the spring 2026 update; ❓ other sizes and intervals | ✅ as 8K, one frame every 2 s, IMU kept, no proxy (Slow motion and timelapse) | 1 at 8K, 2 s, H.265 (a hover) | experimental: stabilization on the flight time (by the tools); Media information shows Timelapse; wanted, low: another interval, one in flight (for the GPS) |
| 360 Photo: Standard, HDR, Burst, AEB | 55 MP 10496×5248, 14 MP 5248×2624 | ✅ .insp, ✅ .dng per AEB bracket (Photos) |
4 at 55 MP, 1 at 14 MP, 1 AEB set (3 DNGs) | not supported (no photos in the app yet); wanted, low: HDR |
| Internal storage (20 GB) vs microSD | ✅ microSD; internal ❓ | all from microSD | wanted, low: a clip from the internal storage | |
| Audio | not in the spec table | ✅ no audio track in our clips | 17 | done: opens and exports as picture only |
| Drone in the picture (props, arms, landing gear) | 3 in flight; _024 take-off, _025 landing |
done in flight: a part at lens 0's rim only, outside the stitched picture; on the ground the landing legs are in lens 0 (Streams), ❓ in the stitch | ||
| Replaceable lenses | none | wanted, low: whether the calibration in the file follows a lens swap | ||
| IMU axes | _010 (flight turns) |
done (-x,y,-z) |
Cinematic, Normal and Sport are flight modes, not recording modes; they change how the drone moves, which matters only for the stabilization check.
Still wanted, 10–30 s each, straight from the card with the whole folder
and with the drone's own file names, as in "Sample clips wanted"
(README.md):
- The
.insvof the third set's_026(only its proxy came). - 4K at 30, 25 or 24 (not Slow Motion).
- An HDR photo, low.
- A clip recorded to the internal storage, low.
Sources
- Shotkit, Antigravity A1 review (specs, formats): https://shotkit.com/antigravity-a1-review/
- Space.com, Antigravity A1 review (sensors, weight): https://space.com/technology/antigravity-a1-drone-review
- Antigravity's spec table (video and photo modes, codecs, storage), as reprinted by apollomaniacs: https://www.apollomaniacs.com/ipod/howto_ar_drone_basic_antigravity_a1.htm
- No Film School, the spring 2026 update (adds timelapse): https://nofilmschool.com/antigravity-a1-drone-spring-update
- Gyroflow telemetry-parser,
src/insta360/mod.rs(the slotted, descriptor-less layout): https://github.com/AdrianEddy/telemetry-parser