SHARP → Splatapult

Turn any single photo into an interactive 3D scene — in one click.

SHARP → Splatapult chains Apple's SHARP model with the splatapult viewer so you can go from a JPEG to a fully interactive 3D Gaussian Splat without touching the command line.

Timeline.1.mp4

---

What it does

📷 photo.jpg
   │
   ▼  Step 1 — Apple SHARP  (AI reconstruction)
   │
   ▼  Step 2 — Coordinate transform  (OpenCV → OpenGL axes)
   │
   ▼  Step 3 — 3dgsconverter  (PLY → .spz, ~10× smaller)
   │
   ▼  splatapult.exe  (real-time viewer, free-fly camera)

---

Requirements

Requirement	Notes
Windows 10 / 11 x64	Required
NVIDIA GPU (GTX 10xx or newer)	Strongly recommended – GPU runs SHARP ~10× faster than CPU
Anaconda or Miniconda	Required
Git for Windows	Required
~6 GB free disk space	For the conda env + PyTorch
Internet connection (first setup only)	To download SHARP weights (~500 MB)

---

Quick Start — Pre-built release (recommended)

1. Download the latest release zip from the Releases page
2. Extract it anywhere (e.g. C:\Tools\SHARP-to-Splatapult\)
3. Run Setup_NewPC.bat once — this installs the sharp conda environment, PyTorch, and all dependencies (~15 min first time)
4. Done! Either:
   • Double-click SHARP_to_Splatapult.exe and pick an image
   • Right-click any JPEG/PNG → Send To → SHARP to Splatapult (shortcut created by Setup)

First run note: SHARP downloads its model weights (~500 MB) automatically on the very first conversion. Subsequent runs are instant.

---

Quick Start — Run from source

git clone https://github.com/Enndee/SHARP-to-Splatapult.git
cd SHARP-to-Splatapult
Setup_NewPC.bat

Then either:

• Run python SHARP_to_Splatapult.py directly from the sharp conda env
• Or build the standalone exe with build_exe.bat

You will also need a compiled splatapult\ folder next to the script.
See Building splatapult below.

---

Release folder layout

After extracting a release zip, the folder looks like this:

SHARP-to-Splatapult\
├── SHARP_to_Splatapult.exe   ← Main GUI (double-click to launch)
├── Setup_NewPC.bat           ← Run once on a new PC
├── splatapult\               ← Viewer binaries (required, do not move)
│   ├── splatapult.exe
│   ├── SDL2.dll
│   ├── glew32.dll
│   └── shader\, data\, ...
└── README.md

The .exe and the splatapult\ folder must stay in the same directory.

---

GUI overview

Element	Function
Click zone / Send To	Select the input image
CUDA / CPU radio	Choose GPU (fast) or CPU (slow) compute
Log area	Live output from SHARP and the converter
Progress bar	Spins while a conversion is running

---

Using the batch script (SHARP_to_Splatapult.bat)

If you prefer not to use the GUI, you can drag-and-drop an image onto SHARP_to_Splatapult.bat directly, or call it from the command line:

SHARP_to_Splatapult.bat "C:\Photos\my_room.jpg"

The .spz output is saved next to the input image.

---

Building splatapult

The viewer is a fork of hyperlogic/splatapult with added .spz loading support (see PR).

git clone --recurse-submodules https://github.com/Enndee/splatapult.git
cd splatapult
git checkout feature/spz-format-support
cmake -B build -S . -DCMAKE_TOOLCHAIN_FILE=vcpkg/scripts/buildsystems/vcpkg.cmake
cmake --build build --config Release

Then copy the build\Release\ contents into a splatapult\ folder next to SHARP_to_Splatapult.exe.

---

Building the .exe from source

build_exe.bat

Requires the sharp conda environment (run Setup_NewPC.bat first). Produces dist\SHARP_to_Splatapult.exe using PyInstaller.

---

How it works

Step 1 — SHARP reconstruction

Apple's SHARP (Single image High-quality Auto-Regressive Predictor) is a transformer-based model that predicts a full 3D Gaussian Splat scene from a single RGB image. It runs via the sharp predict CLI.

Step 2 — Coordinate system fix

SHARP outputs PLY files in OpenCV convention (+Z forward, Y-down). Splatapult uses OpenGL convention (−Z forward, Y-up). The Y and Z position axes and the corresponding quaternion components (rot_2, rot_3) are negated.

Step 3 — SPZ compression

3dgsconverter converts the PLY to Niantic's .spz format — a gzip-compressed binary that is typically ~10× smaller than the equivalent PLY. The PLY is deleted after successful conversion.

Step 4 — Viewer

splatapult renders the .spz file in real time using a free-fly camera. Our fork adds native .spz loading support alongside the existing .ply support.

---

Troubleshooting

"splatapult.exe not found"
Make sure the splatapult\ folder is next to the .exe. If building from source, run the cmake build first.

"Conda env 'sharp' not found"
Run Setup_NewPC.bat. If conda itself is not found, install Anaconda first.

SHARP fails with a CUDA error
Switch to CPU mode in the GUI. Also check your GPU driver is up to date.

SPZ conversion produces a warning and falls back to PLY
Make sure 3dgsconverter is installed: conda run -n sharp pip install git+https://github.com/francescofugazzi/3dgsconverter.git

First run is very slow
SHARP downloads ~500 MB of model weights on the first conversion. This is normal and only happens once.

---

Acknowledgements

Project	License	Role
Apple ml-sharp	SHARP Research License	3D reconstruction model
hyperlogic/splatapult	MIT	Real-time Gaussian Splat viewer
nianticlabs/spz	MIT	SPZ format C++ library
francescofugazzi/3dgsconverter	MIT	PLY → SPZ conversion

---

License

MIT — see LICENSE