Skip to content
360 Gaussian Pro logo360 Gaussian Pro

Alignment

Every alignment setting explained — matcher, lens priors, GLOMAP, feature quality, camera filter and more.

Alignment runs on a custom build of COLMAP 4.2.1. Features are found and matched with the app's own Vulkan SIFT, so this step runs on NVIDIA, AMD and Intel GPUs alike. For the short version, see the First Run walkthrough.

Two stages

Alignment runs in two stages. Stage 1 solves a camera pose for every fisheye frame. Stage 2 finds and matches features on the pinhole tiles those poses produce, and that match graph is what training starts from.

Two-stage alignmentFisheye poses first, then features on the splits
Stage
≈ 3 400 features per fisheye frame · camera positions solved

Matcher

The matcher decides which image pairs COLMAP compares when searching for overlap. Pick it based on how the footage was captured:

  • Sequential — compares each frame with its neighbours in capture order. Best for video, where consecutive frames overlap. Fast, and the right choice for most 360 clips.
  • Exhaustive — compares every image against every other. Highest quality and the most robust for unordered photo sets, but the slowest, since the pair count grows with the square of the image count.
  • Spatial — uses GPS to compare only images that were physically close. Fast and good quality when the footage has reliable GPS, and it also scales the splat to real-world size (with .osv files this is still experimental).
MatcherWhich image pairs COLMAP compares
480 pairs compared for 240 frames

Sequential overlap

Overlap sets how many neighbours on each side of a frame the sequential matcher compares it with. Lower is faster; higher catches more of the scene when the camera moves quickly.

Sequential overlapHow many neighbours each frame is compared with
Overlap5
1 200 pairs · 5 neighbours each side

Loop detection

Sequential matching also checks whether the camera has come back to a place it has already seen. Recognising the same place closes the loop and removes the drift that otherwise builds up along a walk.

Loop detectionReturning to a place you have already seen
loop closed · drift 0.02 m

Lens priors

The app learns your camera's lens distortion on the first alignment and remembers it. In Auto mode it tries, in order:

  1. Your camera's cached calibration — saved after the first successful alignment. Your camera's actual optics, so the best results.
  2. Shipped presets — built-in lens data for known cameras (Insta360 X4/X5/X6, DJI Osmo 360, Avata 360, Antigravity A1), matched from the video's metadata.
  3. Self-calibration — COLMAP estimates the distortion during alignment. Slower, but it works — and the result is cached for next time.
Setting What it does
Auto (recommended) Cache → preset → self-calibrate. Set and forget.
Camera model Force a shipped preset (e.g. "Insta360 X5").
Custom Enter lens parameters by hand. Advanced users only.
Off Always self-calibrate. Slower — geometry appears after 10+ min instead of instantly.

Global mapper (GLOMAP)

Default: ON. Solves all camera positions at once instead of registering frames one by one — 10–100× faster.

Global mapperOne camera at a time, or all of them at once
0 / 14 registered · ≈ 0 min

Leave it on. The first clip from a new camera automatically uses the slower incremental mapper to learn the lens, then all later clips use GLOMAP. Turn it off only when troubleshooting geometry issues, to rule out the solver.

Reuse fisheye features

Default: OFF. Reuses the features already found on the fisheye images (Stage 1) for the pinhole views, instead of detecting new features from scratch on the tiles. Skipping tile feature detection makes the pinhole stage faster — how much faster depends on the clip — and uses much less VRAM.

Trade-off: fine details come out slightly softer, and some can be missing.

It can also change the alignment, because the two paths use different feature caps. Reusing fisheye features draws on the larger fisheye SIFT extraction max of 32,768 features, while detecting fresh on the pinhole tiles is capped at 8,192. Both caps can be changed in the advanced settings.

Reuse fisheye featuresStage 2 skips detection and takes stage 1's features
Pinhole stagefaster
GPU VRAM~1 GB
Point cloudfewer · softer

Does not apply to flat (non-360) input.

faster pinhole stage · varies by clip
Reuse ON Reuse OFF
Pinhole stage speed Faster (varies by clip) Baseline
Point cloud Fewer points; fine detail slightly softer Denser, crisper fine detail
GPU VRAM ~1 GB ~4–8 GB
SIFT extraction max 32,768 (fisheye) 8,192 (pinhole tiles)

Ignored for flat (non-360) input.

Geometry timeline check (experimental)

Fixes "impossible" folded geometry: when two places in your scene look identical (the same wallpaper in two hotel rooms, a stairwell walked up and then down, mirrored floor layouts), the matcher can merge them into one and the model folds onto itself.

The check enforces a simple rule: frames far apart in time must also be far apart in space — otherwise the camera would have to teleport. When a fold is detected, it retries alignment with progressively tighter temporal limits until the fold is gone, and uses GPS (when available) to make sure real loop closures are kept.

Turn it on for interiors with repeating patterns, identical facades or stairwell walks. It only acts if a fold is actually detected.

Result message Meaning
Clean No fold detected — nothing was changed.
Fold detected → cleared at gate W=X The fold was found and fixed automatically.
Fold detected → no gate worked Couldn't fix automatically — mask the duplicated content, split the clip into sections, or trim the problem part.

Feature quality preset

Controls how aggressively Stage 2 extracts and matches features:

Preset Best for Speed (8K) Point density GPU VRAM
Low Low-VRAM cards (~4 GB) ~2–5 min Sparse Minimal
Default Most users ~5–10 min Moderate Standard
High Maximum detail ~15–30 min Dense ~8+ GB

Start with Default. Switch to Low if the GPU runs out of memory during Stage 2, or to High when walls look too sparse after training and you have the VRAM and time — it extracts roughly 30–50% more features. Tweaking any individual setting switches the preset to Custom, which you can save under its own name.

Detail photos

Detail photos and drone or phone videos are added through the batch, not on this tab. Drop them into a batch box, or onto a single 360 clip, in the queue. See Batches.

After Align, the detail images are registered into the finished 360 model automatically. The 360 geometry stays frozen, and the detail adds high-resolution coverage where the 360 video is coarse. The log reports the result as [Detail done] Registered N/M …. Detail material added to an already-aligned project is registered right away.

The model without detail is kept in pipeline/sparse_no_detail/, so you can compare the two.

Collision mesh

Tick Build collision mesh in the Pipeline column to also export a simple surface mesh of the scene. It sits in the same position and scale as the splat — upright and in metres when GPS alignment ran — so you can drop both into a game engine and walk on the floors and bump into the walls.

The mesh is built from the alignment's point cloud, not from the splat. It is coarse and untextured: made for collision and physics, not for looking at. It adds a few seconds to a few minutes at the end of alignment, and if it fails the alignment still finishes.

Pick a preset next to the checkbox:

Preset Best for
INT (default) Interiors. A room is already a closed shell, so only the worst false surfaces are cut and doorways stay open.
EXT Exteriors. Open sky tends to get domed over, so false surfaces are cut harder.

The presets fill in the Mesh tab under Advanced settings, where you can fine-tune detail, gap closing, smoothing and face limit. Changing any value switches the preset to Custom, which you can save by name.

Format Notes
GLB (default) Single binary file — the one to use for Unreal, Unity and Blender.
glTF Self-contained text file.
PLY Opens in MeshLab, CloudCompare and similar.
OBJ Widely supported, but the largest of the four.

The mesh is saved to <project>/mesh/mesh.glb. To try different settings without aligning again, open Advanced settings and press Run Mesh. It rebuilds the mesh from the alignment already on disk. The mesh is also rebuilt automatically when detail photos or stills are added to the model.

If the model is too sparse, no surface comes out. Align more frames, or lower the Stage 2 cleaning.

Export .txt model

Default: OFF. The finished alignment is always saved as COLMAP binary files, which Brush, LichtFeld Studio and Postshot all read. Tick Export .txt model to also write cameras.txt, images.txt and points3D.txt to 0_train_data/sparse/.

Turn it on when you want to open the alignment in RealityScan, which only imports COLMAP text files — for example to build a detailed mesh there. Leave it off otherwise: on big clips the text files take a few minutes and several GB.

Camera filter

Align with every frame, then train with only a share of them. The Camera filter slider sits under the 3D preview at the top of the Alignment tab and runs from 10% to 100% (default — every frame is trained).

Camera filterAlign with every frame, train with a share
Camera filter50%
18 of 36 frames trained · point cloud uses all 36
  • Frames are dropped evenly along the clip: at 50%, every second frame is kept. Both lenses and all splits of one moment stay together, and detail photos and stills are never dropped.
  • Green cameras in the preview are trained; red ones are left out.
  • Alignment and the point cloud still use every frame, so the geometry doesn't get worse.
  • The filter is applied when you press Apply or start training. The left-out images move to pipeline/train_parked/, and a new alignment brings them all back.

At 50%, the splat typically loses 0–0.4 dB (PSNR) — hardly visible. Below 50%, quality drops faster, about 0.8 dB per step. Fewer images train faster with LichtFeld Studio's Auto step count. Brush and Postshot keep the number of steps you set, so lower that too if you want a shorter run.