OpenSplat3DTrainer

Turn a single image into a 3D Gaussian Splatting (3DGS) model — fully local GPU inference, no Python runtime.

中文版 README：README-zh.md

OpenSplat3DTrainer is a local 3DGS generation tool: feed it one image, and in tens of seconds it produces a 3D Gaussian Splatting model you can inspect, import, and export in the built-in 3D viewport.

The inference core is a LibTorch C++ (TorchScript) reimplementation of the TripoSplat image-to-3D method — no Python runtime. The frontend is a WebView2 desktop host serving a local HTTP app; the viewport uses the Spark Gaussian Splatting renderer.

---

Features

• Single-image 3DGS generation — one photo to a high-quality 3D gaussian model
• Batch queue — generate many images in a row, each task with its own parameters
• Full parameter control — sampling steps, guidance, shift, seed, gaussian count, input resolution, matting resolution
• Adjustable input resolution — 512 / 768 / 1024 / 1536, trade speed vs detail
• Real-time 3D viewport — Spark renderer with front / left / top / perspective views, grid, reset, flip, PLY import
• Adjustable rendering — tune smoothing and supersampling live in the "Render" panel
• Dual VRAM modes — resident (performance) vs on-demand (memory-saving)
• Task progress bar — live progress across matting / encoding / sampling steps / decoding
• File logging — log.txt records everything for easy debugging
• Output formats — PLY / SPLAT

---

Rendering comparison

---

Downloads

• GitHub Releases — grab the latest build (exe + runtime libraries + frontend) from the Releases page;
• China cloud drive — mirror links are provided in each release description for faster access in mainland China;
• Model weights (models/, ~3 GB) ship with releases or have a separate download link in the release notes.

---

Hardware requirements

Item	Requirement
GPU	NVIDIA (CUDA 12+)
VRAM	12 GB or more recommended
Minimum	8 GB (low-VRAM mode, on-demand model loading)

• ≥ 10 GB free VRAM: use High VRAM mode (models resident, fastest).
• < 10 GB: use Low VRAM mode (models loaded/released per stage, lower peak).

The UI still opens without an NVIDIA GPU, but generation is unavailable.

---

Quick start

1. Put the six model weights into models/ (exported from TripoSplat): rmbg.pt, dinov3.pt, vae_encoder.pt, flow_model.pt, octree.pt, gs_decoder.pt
2. Deploy OpenSplat3DTrainer_Launcher.exe and the DLLs from output/ to the target machine
3. Launch:
   • First run: pick the VRAM mode (High / Low)
   • "Add images" or drag images straight into the window
   • Select a task to tune its parameters (or batch-generate with defaults)
   • Hit "Start", watch the progress bar
4. Double-click a finished task to view it in the 3D viewport, or import any PLY

---

Inference pipeline

input image
   │  BiRefNet matting (adjustable resolution)
   ▼
feature encoding (DINOv3 + Flux2VAE)
   ▼
flow sampling (steps / guidance / shift) ── per-step progress
   ▼
octree + GS decode → gaussian attributes (pos / color / scale / rot)
   ▼
export PLY / SPLAT + viewport preview

---

Building

• Dependencies: Visual Studio 2022+, CMake 3.20+, LibTorch (CUDA, via CMAKE_PREFIX_PATH)
• WebView2 SDK: not MIT-licensed, so it is not committed; scripts/build.ps1 auto-runs scripts/fetch_webview2.ps1 to pull it from NuGet when missing (or run it manually)
• Core (app/): inference engine + exporters (LibTorch C++)
• Launcher (launcher/): WebView2 desktop host + frontend (HTML/CSS/JS + Three.js + Spark)

# example (point CMAKE_PREFIX_PATH at your torch install)
cmake -S launcher -B launcher/build -DCMAKE_BUILD_TYPE=Release -DCMAKE_PREFIX_PATH="<torch-path>"
cmake --build launcher/build --config Release

Build artifacts deploy to output/ (exe + runtime DLLs + web/ frontend + models/).

Third-party components & licenses

Component	License	How it ships
Spark	MIT	committed (launcher/third_party/Spark, launcher/web/vendor/Spark)
three.js	MIT	committed (launcher/web/vendor/three)
nlohmann/json	MIT	committed (launcher/third_party/nlohmann)
WebView2 SDK	Microsoft permissive (not MIT)	not committed; fetched from NuGet by the build script
TripoSplat model weights	per TripoSplat upstream	not committed (models/); shipped with releases

---

Roadmap

• Image generation models — generate the input image from text / reference images, removing the input-photo dependency;
• Multi-view consistency — strengthen reconstruction quality with multi-angle generation;
• More export formats and rendering backends.

---

Credits

• Tripo / TripoSplat — generation method and model weights;
• antimatter15 / Spark — real-time gaussian splatting renderer;
• three.js — WebGL 3D foundation.

---

Contributing

Contributions of any kind are welcome: bug fixes, feature requests, docs, UI polish.

• Open an issue with a description (attach output/log.txt if relevant);
• Fork and submit a pull request;
• C++ / CUDA core, frontend (HTML / JS / design), or rendering pipeline — pick your lane.

Before contributing, make sure your change at least: builds, doesn't raise the VRAM peak, and keeps the UI/logs in Chinese consistent.

---

License

This project is open-sourced under the MIT License (see LICENSE).

Model weights and the generation method originate from TripoSplat; the renderer and frontend dependencies (Spark, three.js, etc.) follow their respective upstream licenses.