从 WorldMesh 出发，梳理当前最优秀的开源 3D 场景生成项目。
文本/图像输入 → 可导航的 3D 世界。源码已本地保存，待 GPU 后逐一复现。
From WorldMesh to the best open-source 3D scene generation projects.
Text/Image → Navigable 3D Worlds. All source code saved locally, ready for GPU.
WorldMesh 提出了几何优先的 3D 场景生成思路，但代码未开源。我们找到了 7 个可复现的替代方案。 The geometry-first approach that sparked this research. Code not yet released — here are 7 reproducible alternatives.
核心思路：几何优先（Geometry-First）— 文本 → 平面图 → 3D 网格支架（墙面、地面、结构）→ 基于网格条件的图像扩散合成外观 → 3D Gaussian Splatting 输出可导航场景。 支持大规模多房间生成，古罗马到赛博朋克多种风格。用户偏好测试 96.2% 优于基线。 Core Idea: Geometry-First — Text → Floor Plan → 3D Mesh Scaffold (walls, floors, structure) → Mesh-Conditioned Image Diffusion for appearance → 3D Gaussian Splatting for navigable output. Supports large-scale multi-room generation across styles (Ancient Roman to Cyberpunk). 96.2% user preference over baselines.
代码完整发布，社区验证，文档齐全 Complete code, community-verified, well-documented
文本/图像 → 可导航 3D 场景。基于全景生成 + Gaussian Splatting，pip 安装即用。 支持低显存模式（10GB），是目前上手门槛最低的方案。 Text/Image → navigable 3D scene. Panorama generation + Gaussian Splatting, pip install ready. Low VRAM mode (10GB) makes it the lowest barrier option available.
该领域的奠基之作。文本 → 带纹理的 3D 房间网格。迭代生成视图、修复、对齐深度、融合网格。 代码干净稳定，复现性经过大量验证。 The foundational work. Text → textured 3D room mesh. Iterative view generation, inpainting, depth alignment, and mesh fusion. Clean, stable codebase with proven reproducibility.
普林斯顿出品，程序化生成照片级室内外 3D 场景。100% 程序化，无需外部素材。 最成熟的项目（3,214 commits），Mac CPU 也能跑。适合生成训练数据集。 Princeton's procedural generation framework. Photorealistic indoor + outdoor 3D scenes, 100% procedural, no external assets. Most mature project (3,214 commits). CPU compatible — works on Mac without NVIDIA GPU.
最接近 WorldMesh 的方案。单张图像 → 连通的可导航 3D 场景，基于 Fast Layered Gaussian Surfels (FLAGS)。 支持浏览器交互式导航，每个新视角 <10 秒。 Closest to WorldMesh. Single image → connected navigable 3D scenes via Fast Layered Gaussian Surfels (FLAGS). Browser-based interactive navigation, <10s per new view.
代码可用，构建流程较复杂或硬件要求更高 Code available, more complex setup or higher hardware requirements
文本 → 分层 360° 全景 3D 场景 + Gaussian Splatting。沉浸感最强。 需编译 C++ 扩展（Ceres solver、360monodepth），构建过程较复杂。 Text → layered 360° panoramic 3D scene + Gaussian Splatting. Most immersive experience. Requires C++ extension compilation (Ceres solver, 360monodepth).
文本 → 3D 场景（Gaussian Splatting），用户偏好测试 88-95%。 提供预生成输出可跳过耗时阶段，Stage 2 训练约需数小时。 Text → 3D scene via Gaussian Splatting. 88-95% user preference in studies. Pre-generated outputs available to skip slow stages. Stage 2 training takes several hours.
文本 + 空间布局 → 多房间公寓 3D 场景（NeRF）。支持复杂平面图。 多房间能力最接近 WorldMesh，但需要 2+ GPU。 Text + spatial layout → multi-room apartment 3D scenes via NeRF. Supports complex floor plans. Closest to WorldMesh's multi-room capability, but requires 2+ GPUs.
根据你的场景和硬件选择最合适的项目 Pick the right project based on your scenario and hardware
pip 安装，最低 10GB 显存，文本直接出 3D 场景pip install, 10GB VRAM minimum, text directly to 3D scene
WorldGenICCV 2023 奠基之作，代码干净，论文被引最多ICCV 2023 foundational work, clean code, most cited
Text2RoomCPU 也能跑，基于 Blender 程序化生成，Mac 可用CPU compatible, Blender-based procedural generation, works on Mac
Infinigen交互式导航，连通多房间，但需要 48GB 显存Interactive navigation, connected multi-room, needs 48GB VRAM
WonderWorldSIGGRAPH 级别质量，360 度环绕 3D 场景SIGGRAPH-level quality, 360-degree surround 3D scenes
LayerPano3D支持平面图控制，最精确的空间布局方案Floor plan controlled, most precise spatial layout
SceneCraft
所有源码已 clone 至 repos/ 目录，待 GPU 后逐一复现
All repos cloned to repos/ — ready for GPU
| 项目Project | 本地路径Local Path | Stars | 会议Venue | GPU | 状态Status |
|---|---|---|---|---|---|
| WorldGen | repos/WorldGen/ |
1,592 | Independent | 10-24GB | 已保存Saved |
| Text2Room | repos/text2room/ |
1,082 | ICCV 2023 | 16-24GB | 已保存Saved |
| Infinigen | repos/infinigen/ |
6,878 | CVPR 23+24 | CPU OK | 已保存Saved |
| WonderWorld | repos/WonderWorld/ |
717 | CVPR 2025 | 48GB | 已保存Saved |
| LayerPano3D | repos/LayerPano3D/ |
315 | SIGGRAPH 25 | 16-24GB | 已保存Saved |
| RealmDreamer | repos/realmdreamer/ |
297 | 3DV 2025 | CUDA 11.8 | 已保存Saved |
| SceneCraft | repos/SceneCraft/ |
233 | NeurIPS 24 | 2+ GPU | 已保存Saved |
3D 场景生成领域近年关键节点 Key milestones in 3D scene generation
进一步学习和追踪 Further reading and tracking