gsply 0.2.16

Ultra-fast Gaussian Splatting PLY I/O library - pure Python with NumPy and Numba. Features 11x faster GPU transfers and automatic format state tracking for seamless conversions.

gsply

Ultra-Fast Gaussian Splatting PLY I/O Library

93M Gaussians/sec read | 57M Gaussians/sec write | Auto-optimized

Quick Start • Installation • Examples • API Reference • Performance

---

Quick Start

from gsply import plyread, GSData, GSTensor

# Read PLY file (auto-detects format, zero-copy)
data = plyread("model.ply")  # Functional API

# Or use object-oriented API
data = GSData.load("model.ply")  # Classmethod

# Access fields
positions = data.means    # (N, 3) xyz coordinates
colors = data.sh0         # (N, 3) RGB colors
scales = data.scales      # (N, 3) scale parameters
rotations = data.quats    # (N, 4) quaternions

# Save PLY file
data.save("output.ply")  # Uncompressed
data.save("output.ply", compressed=True)  # Compressed (71-74% smaller)

# GPU acceleration (optional)
gstensor = GSTensor.load("model.ply", device='cuda')
gstensor.save("output.compressed.ply")  # GPU compression

Performance: 93M Gaussians/sec read, 57M Gaussians/sec write (400K Gaussians in 6-7ms)

---

Overview

Ultra-fast Gaussian Splatting PLY I/O for Python. Zero-copy reads, auto-optimized writes, optional GPU acceleration.

Key Features

• Ultra-Fast: 93M Gaussians/sec read, 57M Gaussians/sec write
• Zero-Copy: Reads use memory views for maximum performance
• Auto-Optimized: Writes are 2.6-2.8x faster automatically
• Format Support: Uncompressed PLY + PlayCanvas compressed (71-74% smaller) + SOG format
• GPU Ready: Optional PyTorch integration with GSTensor (11x faster transfers)
• Pure Python: NumPy + Numba (no C++ compilation required)
• Object-Oriented API: data.save(), GSData.load(), gstensor.save(), GSTensor.load()
• Format Conversion: normalize(), denormalize() with fused kernels (~8-15x faster)
• Color Conversion: to_rgb(), to_sh() for SH ↔ RGB conversion
• Comprehensive: 406 passing tests, full type hints, extensive documentation

---

Installation

Basic Installation

pip install gsply

Core dependencies: NumPy and Numba (automatically installed)

Optional Features

GPU Acceleration (PyTorch):

pip install torch

Enables GSTensor, plyread_gpu(), plywrite_gpu(), and GPU-accelerated format conversions.

SOG Format Support:

pip install gsply[sogs]

Enables sogread() for reading SOG (Splat Ordering Grid) format files.

Full Installation:

pip install gsply[sogs] torch  # GPU + SOG support

Development:

git clone https://github.com/OpsiClear/gsply.git
cd gsply
pip install -e .[dev]  # Includes pytest, ruff, mypy

---

Examples

Basic I/O

from gsply import plyread, plywrite, GSData

# Read PLY file
data = plyread("model.ply")
print(f"Loaded {len(data)} Gaussians")

# Object-oriented API
data = GSData.load("model.ply")  # Auto-detects format
data.save("output.ply")  # Uncompressed
data.save("output.ply", compressed=True)  # Compressed

# Unpack to individual arrays
means, scales, quats, opacities, sh0, shN = data.unpack()

# Write with individual arrays
plywrite("output.ply", means, scales, quats, opacities, sh0, shN)

Format Conversion

from gsply import GSData
import numpy as np

# Load PLY file (log-scales, logit-opacities)
data = GSData.load("scene.ply")

# Convert to linear format for easier manipulation
data.denormalize()  # Uses fused kernel (~8-15x faster)

# Modify in linear space
data.opacities = np.clip(data.opacities * 1.2, 0, 1)
data.scales *= 1.1

# Convert back to PLY format before saving
data.normalize()  # Uses fused kernel (~8-15x faster)
data.save("modified.ply")

GPU Acceleration

from gsply import GSTensor, plyread_gpu, plywrite_gpu

# Direct GPU I/O (4-5x faster than CPU decompress + GPU transfer)
gstensor = plyread_gpu("model.compressed.ply", device='cuda')

# Or convert from CPU
data = GSData.load("model.ply")
gstensor = GSTensor.from_gsdata(data, device='cuda')

# Access GPU tensors
positions_gpu = gstensor.means  # torch.Tensor on GPU
colors_gpu = gstensor.sh0        # torch.Tensor on GPU

# Filter on GPU
high_opacity = gstensor[gstensor.opacities > 0.5]

# GPU format conversion
gstensor.denormalize()  # GPU-accelerated

# Write back (GPU compression)
gstensor.save("output.compressed.ply")  # GPU compression by default

Creating from External Data

from gsply import GSData, GSTensor
import numpy as np
import torch

# Create GSData from arrays with format preset
data = GSData.from_arrays(
    means=np.random.randn(1000, 3),
    scales=np.random.rand(1000, 3),
    quats=np.random.randn(1000, 4),
    opacities=np.random.rand(1000),
    sh0=np.random.randn(1000, 3),
    format="linear"  # or "auto", "ply", "rasterizer"
)

# Create GSTensor from tensors
gstensor = GSTensor.from_arrays(
    means=torch.randn(1000, 3),
    scales=torch.rand(1000, 3),
    quats=torch.randn(1000, 4),
    opacities=torch.rand(1000),
    sh0=torch.randn(1000, 3),
    format="linear",
    device="cuda"
)

Data Manipulation

from gsply import GSData

# Slicing and indexing
subset = data[100:200]                    # Slice
first = data[0]                          # Single Gaussian
filtered = data[data.opacities > 0.5]    # Boolean mask

# Concatenation
combined = data1 + data2                 # Pairwise (1.9x faster)
merged = GSData.concatenate([data1, data2, data3])  # Bulk (6.15x faster)

# Optimize for faster operations
data = data.make_contiguous()            # 2-45x speedup for operations

# Copy and modify
bright = data.copy()
bright.sh0 *= 1.5  # Make brighter

In-Memory Compression

from gsply import compress_to_bytes, decompress_from_bytes

# Compress for network transfer or storage
compressed_bytes = compress_to_bytes(data)

# Decompress from bytes
data_restored = decompress_from_bytes(compressed_bytes)

SOG Format (Optional)

from gsply import sogread

# Read SOG file - returns GSData (same API as plyread)
data = sogread("model.sog")
positions = data.means
colors = data.sh0

# Read from bytes (in-memory, no disk I/O)
with open("model.sog", "rb") as f:
    sog_bytes = f.read()
data = sogread(sog_bytes)  # Fully in-memory extraction

---

Performance

Benchmark Summary

Uncompressed Format (400K Gaussians, SH0):

• Read: 5.7ms (70M Gaussians/sec)
• Write: 19.3ms (21M Gaussians/sec)

Compressed Format (400K Gaussians, SH0):

• Read: 8.5ms (47M Gaussians/sec)
• Write: 15.0ms (27M Gaussians/sec)
• Size reduction: 71-74%

Peak Performance:

• Read: 78M Gaussians/sec (1M Gaussians, SH0, uncompressed)
• Write: 29M Gaussians/sec (100K Gaussians, SH0, compressed)

GPU Transfer (400K Gaussians, RTX 3090 Ti):

• With _base optimization: 1.99ms (11x faster, single tensor transfer)
• Without _base: 22.78ms (CPU copy + transfer)

Format Conversion (Fused Kernels):

• normalize() / denormalize(): ~8-15x faster with parallel Numba kernels
• Single-pass processing reduces memory overhead

See detailed performance benchmarks for more information.

---

Format Support

Uncompressed PLY

Standard binary little-endian PLY format:

SH Degree	Properties	Description
0	14	xyz, f_dc(3), opacity, scales(3), quats(4)
1	23	+ 9 f_rest coefficients
2	38	+ 24 f_rest coefficients
3	59	+ 45 f_rest coefficients

Compressed PLY (PlayCanvas)

Chunk-based quantized format:

• Automatically saves as .compressed.ply when compressed=True
• Compression ratio: 71-74% size reduction
• Compatible with PlayCanvas, SuperSplat, other WebGL viewers
• Parallel compression/decompression with Numba JIT

SOG Format (Splat Ordering Grid) - Optional

WebP-based texture format for web deployment:

• Requires gsply[sogs] installation
• Uses WebP images for efficient storage
• Codebook-based compression for scales and colors
• Compatible with PlayCanvas splat-transform
• Supports both .sog ZIP bundles and folder formats
• Returns GSData container (same API as plyread())
• In-memory ZIP extraction: Can read directly from bytes

---

API Reference

Complete API documentation: docs/API_REFERENCE.md

Core I/O

• plyread(file_path) - Read PLY files (auto-detects format)
• plywrite(file_path, ...) - Write PLY files
• detect_format(file_path) - Detect format and SH degree
• sogread(file_path | bytes) - Read SOG files (optional)

GSData Container

• GSData.load(file_path) - Load from PLY (classmethod)
• data.save(file_path, compressed=False) - Save to PLY
• GSData.from_arrays(...) - Create from arrays with format preset
• GSData.from_dict(...) - Create from dictionary with format preset
• data.normalize() / data.denormalize() - Format conversion (fused kernels, ~8-15x faster)
• data.to_rgb() / data.to_sh() - Color conversion
• Format query: is_scales_ply, is_scales_linear, is_opacities_ply, is_opacities_linear, is_sh0_sh, is_sh0_rgb, is_sh_order_0/1/2/3
• data[index] - Indexing and slicing
• data.unpack() - Unpack to tuple
• data.copy() - Deep copy

Compression APIs

• compress_to_bytes(data) - Compress to bytes
• compress_to_arrays(data) - Compress to arrays
• decompress_from_bytes(bytes) - Decompress from bytes

Utility Functions

• sh2rgb(sh) / rgb2sh(rgb) - Color conversion
• logit(x) / sigmoid(x) - Optimized CPU functions
• apply_pre_activations(data, ...) - Fused activation kernel (~8-15x faster)
• apply_pre_deactivations(data, ...) - Fused deactivation kernel (~8-15x faster)
• SH_C0 - Normalization constant

GPU Support (PyTorch)

• GSTensor.load(file_path, device='cuda') - Load from PLY (classmethod)
• gstensor.save(file_path, compressed=True) - Save with GPU compression
• GSTensor.from_arrays(...) - Create from tensors with format preset
• GSTensor.from_gsdata(data, device='cuda') - Convert to GPU
• gstensor.to_gsdata() - Convert to CPU
• gstensor.normalize() / gstensor.denormalize() - GPU format conversion
• gstensor.to_rgb() / gstensor.to_sh() - GPU color conversion
• Format query: is_scales_ply, is_scales_linear, is_opacities_ply, is_opacities_linear, is_sh0_sh, is_sh0_rgb, is_sh_order_0/1/2/3
• Device management: .to(), .cpu(), .cuda()
• Precision: .half(), .float(), .double()

---

What's New

v0.2.11 - GPU Compression Optimization

• torch.compile() Auto-Optimization: GPU compression now uses torch.compile() when available
  • ~25% faster GPU compression (5.0ms → 4.0ms for 365K Gaussians)
  • Automatic fallback to eager mode if compilation fails
  • Zero configuration required - works transparently
• GPU Rounding Fix: Fixed quaternion quantization to match CPU behavior exactly
• Platform Support: Triton backend on Linux (standard), Windows (requires triton-windows package)

v0.2.9 - Protocol Interfaces & Performance Optimization

• Protocol Interfaces: Type-safe interfaces for format management
  • FormatAware - Protocol for objects that track format state
  • Normalizable - Protocol for objects that support format conversion
  • GaussianContainer - Protocol for objects containing Gaussian splat data
  • Enables structural typing across GSData and GSTensor
• Format Management API: Advanced format control methods
  • format_state property - Returns current format as immutable dictionary
  • copy_format_from(other) - Copy format state from another object
  • with_format(**kwargs) - Create shallow copy with modified format
• Performance Optimizations: Improved efficiency for common operations
  • Removed auto-consolidate overhead in plywrite() - users can call make_contiguous() manually when needed
  • In-place format conversion now default (inplace=True) - reduces memory allocations
  • Better performance for already-contiguous data

v0.2.8 - Format Query Properties

• Format Query Properties: Convenient boolean properties to check current data format
  • is_scales_ply, is_scales_linear - Check scale format
  • is_opacities_ply, is_opacities_linear - Check opacity format
  • is_sh0_sh, is_sh0_rgb - Check color format
  • is_sh_order_0/1/2/3 - Check SH degree
  • Available on both GSData and GSTensor

v0.2.7 - Fused Activation Kernels & Performance Optimization

• Fused Activation Kernels: Ultra-fast format conversion with parallel Numba kernels
  • apply_pre_activations() - Fused kernel for activating scales, opacities, and quaternions (~8-15x faster)
  • apply_pre_deactivations() - Fused kernel for deactivating scales and opacities (~8-15x faster)
  • normalize() and denormalize() now use optimized fused kernels internally
  • Single-pass processing reduces memory overhead and improves cache locality

v0.2.6 - Format Safety & Auto-detection

• Convenience Factory Methods: GSData.from_arrays(), GSData.from_dict(), GSTensor.from_arrays(), GSTensor.from_dict()
• Auto-Format Detection: Smart heuristics automatically detect PLY vs Linear format
• Format Safety: Strict validation prevents mixing incompatible formats
• Format Helpers: create_ply_format(), create_rasterizer_format()

v0.2.5 - SOG Format Support & API Improvements

• SOG Format Support: sogread() - Read SOG (Splat Ordering Grid) format files
• Object-Oriented I/O API: data.save(), GSData.load(), gstensor.save(), GSTensor.load()
• Format Conversion API: normalize(), denormalize() for linear ↔ PLY format conversion
• Color Conversion API: to_rgb(), to_sh() for SH ↔ RGB conversion

Full Changelog • API Reference

---

Development

Setup

# Clone repository
git clone https://github.com/OpsiClear/gsply.git
cd gsply

# Install in development mode
pip install -e .[dev]

# Run tests
pytest tests/ -v

# Run with coverage
pytest tests/ -v --cov=gsply --cov-report=html

Project Structure

gsply/
├── src/gsply/          # Source code
│   ├── gsdata.py       # GSData dataclass
│   ├── reader.py       # PLY reading
│   ├── writer.py       # PLY writing
│   ├── formats.py      # Format detection
│   ├── utils.py        # Utility functions (fused kernels)
│   └── torch/          # PyTorch integration
│       └── gstensor.py # GSTensor GPU dataclass
├── tests/              # Unit tests (365 tests)
├── benchmarks/         # Performance benchmarks
├── docs/               # Documentation
└── pyproject.toml      # Package configuration

Testing

gsply has comprehensive test coverage with 406 passing tests:

# Run all tests
pytest tests/ -v

# Run PyTorch tests (requires torch)
pytest tests/ -v -k "torch or gstensor"

# Run with coverage
pytest tests/ -v --cov=gsply --cov-report=html

Benchmarking

# Install benchmark dependencies
pip install -e .[benchmark]

# Run benchmark
python benchmarks/benchmark.py

---

Documentation

• API Reference - Complete API documentation
• Changelog - Version history and release notes
• Contributing - Contribution guidelines

---

Contributing

Contributions are welcome! Please see docs/CONTRIBUTING.md for guidelines.

Quick start:

1. Fork the repository
2. Create a feature branch
3. Make your changes with tests
4. Run tests: pytest tests/ -v
5. Submit a pull request

---

License

MIT License - see LICENSE file for details.

---

Citation

If you use gsply in your research, please cite:

@software{gsply2024,
  author = {OpsiClear},
  title = {gsply: Ultra-Fast Gaussian Splatting PLY I/O},
  year = {2024},
  url = {https://github.com/OpsiClear/gsply}
}

---

Related Projects

• gsplat: CUDA-accelerated Gaussian Splatting rasterizer
• nerfstudio: NeRF training framework with Gaussian Splatting support
• PlayCanvas SuperSplat: Web-based Gaussian Splatting viewer
• 3D Gaussian Splatting: Original paper and implementation

---

Made with Python and numpy

Report Bug • Request Feature • Documentation