Shrink laser-scanned point clouds without uploading them.

A free, on-device point cloud studio that opens PLY, PCD, and XYZ scans, lets you crop, downsample, recolor, and filter outliers, and writes a clean PLY, XYZ, or GLB point cloud without ever uploading the source. It is built for surveyors, LiDAR and SLAM researchers, photogrammetry hobbyists, drone mappers, and anyone who needs to make a million-point scan a tenth of its size before sharing it.

Five downsamplers, six bounding-box sliders, statistical outlier removal, axis alignment, and an asset report with point counts, bounds, density, and channel coverage — the same day-to-day workflow a desktop tool offers, in a tab.

Five downsamplers, picked for the scan in front of you

Voxel-grid keeps the cloud at a uniform density — every cube of the chosen size contributes at most one point, picked as the centroid of the cube’s contents. It is the right choice for dense LiDAR rasters where you just want fewer points without losing the surface. Random is a quick-and-dirty 1-in-N downsample for previews and share links. Every-Nth is a deterministic stride, useful for time-aligned series where you want a fixed temporal spacing. Color-histogram merges nearby colors before sampling, which keeps vegetation, road, and façade regions intact when a scan was captured with rich RGB. Farthest-point sampling preserves the global shape and is the right choice for sparse SLAM and small object scans.

All five are bit-exact reproducible: the same input file, the same settings, the same output bytes. That matters when you want to compare two parameter sweeps without re-running an entire pipeline.

Cropping, axis alignment, and density maps

Six sliders set a min/max for X, Y, and Z. The 3D preview shows the retained box as a wireframe, so you can dial the crop in while looking at the points it will keep. The 3-axis align-to-Z-up toggle computes the principal direction of the cloud and rotates it so the up vector is +Z — many terrestrial scanners report Y-up, and most web viewers expect Z-up, so this single toggle removes the most common viewing-day headache. Centering on the origin translates the cloud so its bounding-box midpoint lands at 0,0,0; normalizing scales the longest axis to ±1, useful when you want to drop a point cloud into a fixed-scale viewer without doing the math by hand.

The viewer can color the cloud by source RGB, by height (Z), by intensity (when present), or by local density (k-NN distance, normalized). Local density is the right choice to surface under-sampled regions of a SLAM cloud — they show up as bright halos in the density color map.

Statistical outlier removal, and why it runs in one pass

Outliers are the bane of every point cloud pipeline: isolated returns from a LiDAR hitting a raindrop, a SLAM feature matching a window reflection, a photogrammetry tie point on a moving object. The classic k-NN statistical filter computes the mean distance to each point’s k nearest neighbors, then drops every point whose mean distance is more than N standard deviations away from the global mean. It runs in one pass over the cloud, in a typed-array pipeline, so a 10-million-point scan takes under a second on a modern laptop and the memory peak is bounded by the input size plus a small per-point bookkeeping array.

The defaults (k=8, sigma=2.0) are tuned for terrestrial LiDAR; raise k to 20–30 for sparse SLAM clouds where per-point neighborhood statistics are noisier, and raise sigma to 3.0+ if you want to keep more points and only filter the most extreme outliers.

Format bridge: PLY, PCD, XYZ, and GLB points

The parser handles binary little-endian and big-endian PLY (including vertex colors, normals, intensity, and alpha channels), ASCII and binary PCD, and ASCII XYZ with optional RGB columns. It auto-detects the format from the magic bytes and the file extension, so dropping a `.ply`, a `.pcd`, or a `.xyz` file Just Works. PLY properties that are not vertex-x/y/z or vertex-red/green/blue/alpha are preserved in the output if you keep the same format, or silently dropped if you switch to XYZ.

The GLB point cloud export uses the Khronos `KHR_points` extension to pack positions and colors into a single GLB, viewable in three.js, Babylon.js, Blender, and any web 3D viewer. It is the right format to share on a webpage, embed in a design review, or load into a game engine.

Limits, formats, and licensing

Free tier: a single file up to 200 MB or 50 million points, with the full pipeline (voxel, random, every-Nth, color histogram, farthest-point, crop, outlier, color edit, axis alignment, GLB export). Pro: a single file up to 1 GB or 500 million points, and 20 files per batch. A single file is always accepted — batch limits only apply when you queue more than one file. The parser streams the input once into a flat Float32Array, so memory peaks are roughly 2× the source file size; a 200 MB scan is comfortable on any modern laptop, and a 1 GB scan needs roughly 4 GB of free RAM.

Outputs are byte-exact reproducible: the same input and the same settings always yield the same output bytes, so a downstream pipeline (mesh reconstruction, classifier, viewer) sees a stable artifact. PLY outputs use binary little-endian by default, or ASCII with the precision you pick in the advanced panel. GLB exports use KHR_points, so the file is also readable by the broadest set of 3D viewers. Everything is yours, with no watermark, and commercial use is included on both the free and Pro plans.