LibreAI / AI 3D / 3D Gaussian Splatting
// ai-3d · graphdeco-inria/gaussian-splatting

3D Gaussian Splatting

3D Gaussian Splattingは、graphdeco-inriaが公開しているAI 3Dプロジェクトです(GitHub: graphdeco-inria/gaussian-splatting)。GitHubスターは24,460、フォークは3,494で、«Research-only»の条件の下で公開されています。GitHubでの最終更新は2025年10月17日です。

The official SIGGRAPH 2023 implementation — real-time radiance fields at 30+ fps from ordinary photos, research licence

★ 24.5k スター⑂ 3.5k フォークResearch-only更新 2025年10月
▦ ビジュアル — このプロジェクトのREADMEには埋め込み可能なスクリーンショットがないため、
ここではプレースホルダーを表示しています。下のガイド自体は完全な内容です。
// プロジェクトの基本情報

プロジェクト基本情報

GitHubスター
24,460
フォーク
3,494
ライセンス
Research-only
GitHubでの最終更新
2025年10月17日
カテゴリ
AI 3D
掲載タイプ
完全なインストールガイド
公式サイト
Project page
// 仕組み

仕組み

  1. Clone with submodules

    The repo contains submodules, so clone it recursively.

  2. Create the conda environment

    Build the environment from the provided file and activate it.

  3. Prepare your photos

    Use the included script to turn a set of photos into SfM inputs, or start from the sample datasets.

  4. Train, then view in real time

    Train the model, then open the real-time viewer and fly through the scene at 30+ fps.

// 機能

選ばれる理由

Real-time radiance fields

Novel-view synthesis at 30 fps or more in 1080p — the headline result of the paper.

State-of-the-art quality

Matches or beats NeRF-style methods on established datasets while training faster.

Live training viewer

A network viewer lets you watch the optimisation as it happens.

OpenGL real-time viewer

Render trained scenes interactively, with a top view and VR support added since release.

Your own photos in

A helper script converts ordinary photo sets into optimisation-ready SfM data.

// 自分で実行

自分で実行してみる

  1. Step 1 — Clone recursively

    Submodules are required, so use the --recursive flag.

    git clone https://github.com/graphdeco-inria/gaussian-splatting.git --recursive
  2. Step 2 — Create the environment

    Build the conda environment from the repo file.

    conda env create --file environment.yml
    conda activate gaussian_splatting
  3. Step 3 — Train on a scene

    Train from an SfM dataset directory; see the README for the full command and options.

    python train.py -s <path to COLMAP or SfM dataset>
  4. Step 4 — Render the trained scene

    Generate renderings from the trained model with the render script.

    python render.py -m <path to trained model>
// 活用例

作れるもの

  • Rebuild a real place in 3D from a set of photos
  • Render novel views of a scene in real time
  • Learn how radiance fields work from the reference code
  • Benchmark new 3DGS methods against the original

Close neighbours on the AI 3D shelf: Nerfstudio — the plug-and-play studio for NeRFs and Gaussian, and 4d-gaussian-splatting — ICLR 2024] Live Photorealistic Dynamic Scene. The full shelf is the AI 3D tools in the catalogue.

// よくある質問

よくある質問

What is 3D Gaussian Splatting?

A way to rebuild a 3D scene from photos as millions of 3D Gaussians, then render new viewpoints of it in real time — the SIGGRAPH 2023 paper from Inria and MPI.

Can I use it commercially?

Not without permission: the licence allows research and evaluation use only, and commercial use needs prior explicit consent from Inria.

What hardware do I need?

A CUDA-ready GPU (compute capability 7.0+) — 24GB of VRAM for paper-quality training; smaller configs are covered in the README FAQ.

How is it different from NeRF?

Same goal — novel-view synthesis — but much faster: real-time rendering at 1080p instead of seconds per frame.

Where do I start?

Clone with --recursive, create the conda environment, and follow the step-by-step tutorial linked in the README.

// さらに探索

似ているツール

AI 3D

Nerfstudio

Nerfstudio — the plug-and-play studio for NeRFs and Gaussian splatting: turn photos of a real scene into explorable 3D, with a modular API and viewer, Apache-2.0

★ 12.1k スターApache-2.0
ガイドを見る →
AI 3D

4d-gaussian-splatting

ICLR 2024] Live Photorealistic Dynamic Scene Representation and Rendering.

★ 1.0k スターMIT
ガイドを見る →
AI 3D

Gaussian-LIC

Real-time LiDAR-inertial-camera Gaussian Splatting SLAM.

★ 658 スターCustom
ガイドを見る →
AI 3D

GS-SDF

LiDAR-augmented Gaussian Splatting plus neural SDF, in real time.

★ 535 スターGPL-2.0
ガイドを見る →
AI 3D

MILo

Mesh-in-the-loop Gaussian Splatting from a SIGGRAPH Asia paper.

★ 528 スターNOASSERTION
ガイドを見る →
AI 3D

QuestRoomScan

Real-time 3D room reconstruction on Meta Quest 3.

★ 70 スターNOASSERTION
ガイドを見る →
// 代替候補

代替候補: AI 3D

LibreAIカタログに掲載されたAI 3Dプロジェクトの中で最もスターの多いもの — 3D Gaussian Splattingが合わなかった場合に検討したい、現実的な候補リストです。

  • mcp-for-blender★ 30.2k

    Control Blender with natural language: an MCP plugin that lets any LLM model, light and animate in 3D

  • editor★ 24.7k

    Pascal Editor: an open-source 3D architectural editor with MCP tools for humans and AI agents

  • text-to-cad★ 18.4k

    Give your AI agent CAD superpowers - generate real STEP/STL models from text

  • img2threejs★ 17.7k

    Turn a reference image into a procedural, animation-ready Three.js model - as code

  • SolidCraft★ 16.4k

    Open-source parametric 3D CAD in pure Rust — desktop, browser, and an MCP server for AI agents

// さらに見る ai-3d

さらに見る: AI 3D

カタログの中で3D Gaussian Splattingとスター数が近いAI 3Dプロジェクト — 前後の棚もそのまま見て回れます。

  • editor★ 24.7k

    Pascal Editor: an open-source 3D architectural editor with MCP tools for humans and AI agents

  • text-to-cad★ 18.4k

    Give your AI agent CAD superpowers - generate real STEP/STL models from text

  • mcp-for-blender★ 30.2k

    Control Blender with natural language: an MCP plugin that lets any LLM model, light and animate in 3D

  • img2threejs★ 17.7k

    Turn a reference image into a procedural, animation-ready Three.js model - as code

  • NURBGen★ 34

    Text-to-CAD: generate NURBS models from plain language.

  • SolidCraft★ 16.4k

    Open-source parametric 3D CAD in pure Rust — desktop, browser, and an MCP server for AI agents

3D Gaussian Splattingを入手する

ソースコードを見て、リポジトリにスターを付け、作り始めましょう。