3D Gaussian Splatting · Foundations of 3DGS

The Gaussian Splatting License: What You Can and Cannot Do

8 min read

Overview

Before building anything with the original gaussian-splatting codebase, you need to understand what its license allows. The short version: this software is free for research and evaluation only. Commercial use requires explicit prior consent from the licensors.

This reading walks through the license in plain language and highlights the decisions it forces on practitioners.

Who Holds the Rights

The software is owned jointly by Inria (Institut National de Recherche en Informatique et en Automatique, France) and the Max Planck Institut für Informatik (MPII) in Germany. It was registered with the Agence pour la Protection des Programmes (APP) in France.

The licensor's stated goal is to allow the research community to use, test, and evaluate the software — not to enable commercial deployment.

What the License Grants

Section 3 grants:

  • Non-exclusive rights to use the software
  • Free of charge, with no royalty
  • For research and/or evaluation purposes only — both academic and industrial researchers are included in this category
  • The right to reproduce, prepare derivative works, publicly display and perform, and distribute the work and any derivatives

The key phrase is "non-commercially". Industrial research teams can use it internally for evaluation, but cannot deploy it in a product or service without further authorization.

The Commercial Use Restriction

Section 5 is blunt:

THE USER CANNOT USE, EXPLOIT OR DISTRIBUTE THE SOFTWARE FOR COMMERCIAL PURPOSES WITHOUT PRIOR AND EXPLICIT CONSENT OF LICENSORS. YOU MUST CONTACT INRIA FOR ANY UNAUTHORIZED USE: stip-sophia.transfert@inria.fr

This is not a grey area. Deploying the original gaussian-splatting code — or any derivative that inherits this license — in a commercial product, SaaS platform, or client-facing service without written authorization from Inria constitutes a license violation.

What counts as commercial use?

The license does not enumerate specific scenarios, which means when in doubt, contact Inria. Common cases that almost certainly require authorization:

  • Selling or licensing 3DGS-generated assets
  • Offering a reconstruction service where customers pay for output
  • Embedding gaussian-splatting in a commercial application (desktop, web, or mobile)
  • Using it internally to replace a paid tool in a revenue-generating workflow

Derivative Works

Section 4.2 allows derivative works, but with a catch: your derivative must preserve the non-commercial use limitation. You cannot write a wrapper, fork the repo, and redistribute under a permissive license to escape the restriction. Any derivative work must carry the same Section 2 use limitation forward.

This is relevant if you are:

  • Building a Python package on top of gaussian-splatting
  • Creating a training framework that calls into it
  • Releasing modified training scripts

All of these inherit the non-commercial restriction unless you have separately negotiated rights.

Redistribution Requirements

If you redistribute the work (or derivatives), Section 4.1 requires:

  1. You distribute it under this same license
  2. You include a complete copy of the license
  3. You retain all copyright, patent, trademark, and attribution notices without modification

The MIT-Licensed Exception

Section 6 carves out one file: utils/loss_utils.py. This file is based on pytorch-ssim by Evan Su, which is MIT-licensed. The MIT license is permissive — that portion of the code can be used freely in commercial products. The carve-out does not extend to the rest of the repository.

Alternatives for Commercial Use

If you need commercial rights, your options are:

  1. Contact Inria directly at stip-sophia.transfert@inria.fr to negotiate a commercial license
  2. Use an alternative implementation that is permissively licensed. Several community reimplementations (e.g., nerfstudio's splatfacto, gsplat) are MIT or Apache 2.0 licensed and can be used commercially
  3. Build on top of permissively-licensed forks that were developed independently of the Inria codebase

The ecosystem has matured enough that commercial projects are not forced to use the original codebase — but it is worth auditing your dependency tree carefully, since some libraries do link back to components with this restriction.

Key Takeaways

Scenario Allowed?
Academic research Yes
Industrial R&D / evaluation Yes
Open-source non-commercial project Yes, with license included
Commercial SaaS or product No — contact Inria
Fork and redistribute under MIT No
Use utils/loss_utils.py commercially Yes (MIT carve-out)