Tools landscape

Training software split cleanly by 2026 into three tiers: a frozen research reference, a set of permissively licensed open trainers built around a shared rasterisation backend, and commercial desktop apps aimed at production pipelines. Picking between them is now mostly a licensing and workflow question rather than a quality one, since the underlying optimisation loop covered elsewhere in this knowledge base is broadly the same everywhere.

The reference implementation: the baseline, no longer moving

graphdeco-inria/gaussian-splatting, the code released alongside the original 2023 paper, is still the implementation most other tools trace their lineage to, and it’s worth reading even where it isn’t used directly: the iteration counts, loss weighting and densification thresholds described throughout this knowledge base all describe this codebase’s defaults. Two things matter commercially. First, its licence, the custom “Gaussian-Splatting License”, restricts use to research and evaluation: the licence text states the software may be used “non-commercially, i.e., for research and/or evaluation purposes only,” and prohibits any commercial use, exploitation or distribution without prior, explicit consent from the licensors, INRIA and the Max Planck Institut fur Informatik jointly hold the rights, and commercial enquiries go through INRIA’s technology transfer office. Second, the repository’s commit history shows no activity since October 2024; it isn’t formally deprecated, but it has settled into a stable reference rather than a living project, and the ecosystem has moved on to building around it rather than on it.

gsplat: the library other tools build on

gsplat (nerfstudio-project/gsplat) is a CUDA-accelerated rasterisation library with Python bindings, released under Apache-2.0, that reimplements the core rendering and backward pass with claimed lower memory use and higher throughput than the original INRIA CUDA kernels. It’s maintained under the nerfstudio-project organisation and serves as the rasterisation backend behind nerfstudio’s own Gaussian-splatting training method, splatfacto. The practical effect is that gsplat has become the de facto open-source backend that a lot of research code, and several production tools, build training loops on top of rather than each maintaining its own CUDA kernel from scratch.

Brush: cross-platform, no CUDA required

Brush (ArthurBrussee/brush) takes a different technical route: it’s written in Rust on the Burn machine-learning framework with a WebGPU/wgpu backend, deliberately avoiding a hard CUDA dependency. Licensed Apache-2.0, it runs on macOS, Windows, Linux, Android, and in-browser via WebGPU, and ships as dependency-free binaries. For anyone without an NVIDIA GPU in the room, or wanting to run or inspect the actual optimisation loop somewhere a CUDA toolchain isn’t practical, it’s currently the most viable route to training outside a closed commercial app.

PostShot: the commercial desktop option

PostShot, from Jawset, is the commercial end of the spectrum and the tool most production pipelines reach for, including this one. It’s Windows-only, requires an NVIDIA GPU of compute capability 7.5 or newer, and is sold on a subscription with Free, Indie and Studio tiers; the Free tier is explicitly non-commercial, so shipping client work requires a paid tier. Version 1.1 shipped 2026-05-13, following a beta period that ended with the 1.0 release in August 2025. Its feature set is aimed squarely at a production workflow rather than a research one: live preview and editing during training, merging multiple trained scenes, direct video-to-splat ingestion with automatic frame extraction, PLY and SPZ export, and, on the Studio tier, HDR and RAW input plus command-line automation. It’s worth being precise about what isn’t publicly documented: Jawset doesn’t disclose which specific densification strategy or 3DGS variant it trains with, so it’s best treated as a well-supported black box rather than something to reason about at the same level of detail as the open trainers above.

A 2026 addition worth knowing: LichtFeld Studio

LichtFeld Studio (MrNeRF/LichtFeld-Studio) is the clearest new entrant since the tools above became established: an actively developed, GPLv3-licensed desktop app that trains, edits, renders and exports splats from one interface, with a Vulkan viewer, an undo-capable scene editor, a Python plugin system, and export to PLY, SOGCompressed splat format storing Gaussian attributes as sorted 2D image grids (typically WebP), giving large size reductions for web delivery. and a standalone HTML viewer. It requires an NVIDIA GPU with CUDA 12.8 or newer. It’s the closest open-source answer to PostShot’s all-in-one niche, but its GPLv3 licence is a materially different commitment from Apache-2.0: copyleft obligations apply if a product is ever distributed that’s built on top of it, which matters if the plan is ever to embed rather than just run it standalone.

Licence is the deciding question now

Put plainly: graphdeco’s reference code is research-only and needs INRIA’s consent for anything commercial; gsplat and Brush are permissive open source, safe to build on without asking anyone; LichtFeld Studio is open source but copyleft, fine to run, a real constraint to redistribute inside; and PostShot is commercial software where a paid tier buys the right to ship client work.

In practice

For a pipeline that trains client-facing scenes in PostShot, the licence question is already answered by the subscription tier in use, but it’s worth knowing where the alternatives sit before a project ever needs one. Brush or gsplat are the tools to reach for if a scene needs training somewhere without an NVIDIA GPU, or if a task calls for actually inspecting or modifying the optimisation loop rather than trusting a GUI’s defaults. The graphdeco reference is worth keeping open in a second tab for exactly what it’s good for, understanding what a parameter or threshold in this knowledge base actually refers to, and not for anything that will leave the research folder.