Skip to content
360 Gaussian Pro logo360 Gaussian Pro

Extraction

How frames are pulled from your clip, the split count, rolling shutter correction and HDR export.

Extraction turns your video into the still frames the rest of the pipeline works from. For the short version, see the First Run walkthrough.

How it works

Add your clip (for example an Insta360 X5 .insv) to the queue. The app detects the input type — equirectangular, dual-fisheye or flat — and sets up the project automatically.

Frames are then extracted from the video. Rather than sampling at a fixed rate, the app scores each candidate for sharpness and keeps the sharpest frame in each interval, so motion blur and rolling-shutter smear are avoided where possible.

Splits per frame

For 360 input, each extracted frame can be re-projected into 1, 5 or 9 pinhole tiles ("splits"). More splits give the matcher more overlapping, low-distortion views to work with, at the cost of more images to process.

Pinhole splitsOne fisheye lens, re-projected into pinhole views

Equirectangular input is always split into 6 cube-map faces instead.

Frames240
240 frames × 2 lenses × 5 = 2 400 images
Splits Best for
1 Quick tests, or when the scene is small and simple
5 (recommended) The default for almost everything — good coverage, sane processing time
9 Wide, detailed scenes where extra overlap helps alignment

Start with 5 splits. Only move to 9 if alignment struggles to register a large or complex scene, and drop to 1 when you just want a fast preview.

Rolling shutter correction

Default: ON. 360 cameras read their sensor out row by row, so when the camera turns during a frame, straight lines lean and bend. The app straightens every extracted frame using the camera's own motion data, before masking and alignment ever see it.

Rolling shutter correctionThe sensor is read row by row while the camera turns
Turn speed90°/s
the lean grows with turn speed

The setting is in the Extract tab, under Pinhole Split. It works with raw files from these cameras:

Camera File
Insta360 X4 / X5 / X6 .insv
Antigravity A1 .insv
DJI Osmo 360, DJI Avata 360 .osv

The faster you turn, the bigger the gain. On a fast handheld walk with an X6, the trained splat came out about 0.8 dB sharper (PSNR) and alignment error dropped by a fifth. On a slow indoor walk the difference is small. It runs on the CPU, takes well under a second per frame, and needs no calibration.

It's skipped — with a note in the log — for stitched 360 exports, flat video, GoPro clips, batch projects, and files that carry no motion data. Leave it on.

HDR training data

Insta360 X6 clips recorded with the I-Log setting hold 10-bit colour. Tick HDR in the Extract tab, under Pinhole Split, and the app also exports HDR training images: scene-linear, half-float .exr tiles with the same names as the normal ones. Alignment still runs on the normal images; the EXR tiles go along for training.

Setting Options
HDR compression PIZ (lossless) (default) or DWAA (smaller) — about 2–3× smaller files, visually lossless.
Color space ACEScg (default) or Rec.2020.
  • The checkbox is active only for a single I-Log .insv clip, not a batch.
  • HDR training needs Postshot (Studio plan). Ticking HDR switches the Train tab to Postshot, and the app trains Postshot from the EXR tiles.
  • The tiles are written after alignment to 0_train_data_hdr/ in the project folder. They're large — around 15 MB each with PIZ.
  • The trained splat is linear ACEScg (or Rec.2020) with values above 1.0, so view and grade it in an HDR-aware tool.