// 2026.05.18

3D Gaussian Splatting from a 360° Camera: a Capture-to-Scene Workflow

10:00 JST

Photoreal environments are a fast way to make an XR demo feel real. 3D Gaussian Splatting (3DGS) turns ordinary footage into a renderable point-based scene — and a 360° camera makes capture quick. This post summarizes a clear, practical pipeline.

This is a condensed, English walkthrough of @Tks_Yoshinaga’s Qiita article “Getting Started with 360° Image Gaussian Splatting (Basic Workflow)”. Please read the original for full detail and screenshots: qiita.com/Tks_Yoshinaga/items/354e9082bd607f3cefee. All credit to the author; any errors in summarizing are ours.

What you need

  • Camera: a 360° camera — Insta360 X4 Air (shoot 8K) or a RICOH THETA.
  • PC: Windows 11, an NVIDIA RTX GPU (the author used an RTX 4070 SUPER), 32 GB RAM.

The pipeline

1 · Capture & export. Shoot a short clip (start with under one minute). Export to equirectangular with the camera’s software (e.g. Insta360 Studio). Critical: if stabilization is on, disable horizon leveling, tilt recovery, and vibration reduction — they distort the camera geometry SfM relies on.

2 · SfM & point cloud — 360° Gaussian. Load the equirectangular video, set a frame interval, and enable sharp-frame extraction. Optionally mask moving subjects/low-texture regions with AutoMasker (paid) using period-separated keywords like person.sky. Configure: Training Method No Training, SfM SphereSFM, matcher sequential, preset size matching the camera (e.g. Ultra 8K). For large outdoor scenes: more iterations/refinements, lower MinInliers, looser reprojection thresholds. Run, then verify camera poses and the cloud in train_data/sparse.

3 · Train — LichtFeld Studio. Drag the train_data folder in and wait for the point cloud and all cubemap faces to finish loading before training (training mid-load crashes it). Strategy MRNF; Steps Scaler 1 for ≤300 images, else image_count / 300; bump it 2–3× if the view whites out; raise Max Gaussians for more detail. Start training — it sharpens progressively and halts at the step limit.

4 · Export & view. Export a .ply and open it in the browser-based SuperSplat Editor to clean up and render video. Sample result: youtube.com/watch?v=n-NL1UisVF4.

Tools & repositories

Tool Purpose
360° Gaussian Frame extraction + SfM automation (SphereSFM)
AutoMasker Auto-masking of moving/low-texture regions (paid, ~€46)
LichtFeld Studio · source GUI Gaussian Splatting trainer
SuperSplat Editor Browser viewer / cleanup / video export
Insta360 Studio Equirectangular export (native to Insta360)

Gotchas

  • Wait for the full load before training in LichtFeld Studio — the most common crash.
  • Stabilization sub-features must be off at capture or SfM accuracy degrades.
  • First-run “Failed to process” can appear instantly — click Stop and retry.
  • Licensing: prebuilt LichtFeld Studio v0.5.2 is no longer free; v0.4.0 prebuilt remains free, or build v0.5.2 from source. Build tips: git clone --recursive (submodules), and Strawberry Perl can shadow system cmake — rename C:\Strawberry\c\bin\cmake.exe if the build picks the wrong one. See the Windows build wiki.

Why this matters for the hackathon

A captured splat scene is a drop-in photoreal backdrop or asset for a Spectacles/Quest/WebXR build — minutes of capture instead of hours of modeling. The author also has follow-up “quality improvement” and “Metashape” editions worth reading once the basics click.

Questions? Reach us via the Contact page.

// BACK TO NEWS