The 2026 tool landscape by pipeline stage

Nothing in the splat pipeline is one tool end to end. A scene passes through capture, structure-from-motion, training, cleanup, optionally segmentation, compression and a viewer, and a different piece of software, often from a different team under a different licence, owns each stage. This article is a map of that pipeline by stage rather than a repeat of any single article’s detail: each section names the tools this knowledge base has verified, with a one-line liveness and licence judgement, and points at the article covering that stage properly.

Capture

Everything starts as a folder of overlapping photos or a video walk, covered in depth in this knowledge base’s capture part. Scaniverse, the Niantic-owned mobile scanning app, is the one capture-side tool verified elsewhere in this knowledge base: it’s the app SPZ was built around and ships that format natively.

Structure-from-motion

Camera poses and a sparse point cloud come out of COLMAP, the open implementation of the incremental structure-from-motion pipeline described by Schönberger and Frahm, covered fully in this knowledge base’s SfM part. It’s the near-universal choice: essentially every tool downstream of it in this map, PostShot included, expects a COLMAP or COLMAP-shaped dataset as its input.

Training

Covered in the training part’s tools landscape article, the trainer choice splits by licence more than by quality: the graphdeco reference implementation is frozen and research-only; gsplat and Brush are permissively licensed, Apache-2.0, open source, actively maintained, and don’t require an NVIDIA GPU in Brush’s case; LichtFeld Studio is open source but copyleft, GPLv3; and PostShot is the commercial, production-aimed desktop option this pipeline actually trains through.

Editing and cleanup

Covered in the editing and cleanup article in this part: SuperSplat is the default browser-based editor for crop, floater removal and colour, actively developed at v2.28.1; splat-transform is its scriptable, MIT-licensed CLI counterpart for batch and repeatable cleanup; and the KIRI Engine and SplatForge Blender addons extend editing into a full 3D content-creation environment when a splat needs to sit alongside mesh geometry.

Segmentation and semantics

The one stage in this map without a settled tool, covered in the segmentation and semantics article in this part. Gaussian Grouping and SAGA are the two reference research codebases for whole-scene and promptable object segmentation respectively, both usable but neither actively maintained as of mid-2026, and neither integrated into any editor covered above. Treat this stage as custom engineering, not a tool choice, until that changes.

Compression and delivery formats

Covered in the compression-formats part: PLY is the lossless, universal interchange format every trainer writes; .splat is a dependency-free but SH-less shortcut for small demos; SPZ is Niantic’s open, general-purpose compressed format, natively read by Scaniverse, Adobe’s Rotate Object and Babylon.js; and SOG, the self-organising-grid format this pipeline delivers through, is what PlayCanvas, SuperSplat and superspl.at are all built to read natively.

Viewers and engines

Covered in the web-delivery part: PlayCanvas is the only engine with a first-party splat stack end to end, renderer, format and editor built by one team; Babylon.js has genuine first-party support of its own, with native SOG loading added in its March 2026 release; three.js gained a built-in splat addon in r186 (September 2026) that reads PLY, SPZ and glTF but not SOG; and Spark is where the wider three.js ecosystem landed for streamed and LOD scenes, now that both of three.js’s original splat libraries point new users to it directly.

In practice

Six of these seven stages have a clear, licence-driven default in 2026: COLMAP for structure-from-motion, PostShot or an open trainer for training, SuperSplat and splat-transform for cleanup, SOG or SPZ for delivery format, and PlayCanvas or Babylon.js for rendering, depending on which ecosystem a project already lives in. Segmentation is the exception, and it’s worth budgeting it as one: if a brief calls for object-level interaction with a splat, that’s a research integration or a custom detection pipeline, not a line item that any tool above already ships. The general habit that keeps this map useful past its publication date is the same one this knowledge base repeats stage by stage: check a tool’s commit history and licence before adopting it, since this ecosystem has turned over its defaults more than once already.