fast-ngp 0.1.0

Pure PyTorch implementation of NVIDIA's Instant Neural Graphics Primitives

Fast-NGP

A clean, pure PyTorch implementation of NVIDIA's Instant Neural Graphics Primitives with multiresolution hash encoding.

Features

✨ Pure PyTorch - No CUDA extensions required, easy to install and modify
📦 Easy Installation - pip install pyinstant-ngp
🎯 Modular Design - Each component usable independently
🚀 Multiple Tasks - NeRF, SDF, and Gigapixel image support
📊 Visualization Tools - Built-in rendering and metrics
🧪 Well Tested - Comprehensive test suite
📚 Great Docs - Extensive documentation and examples

Installation

From PyPI (recommended)

pip install fast-ngp

From source

git clone https://github.com/Dox45/fast-ngp.git
cd fast-ngp
pip install -e .

Quick Start

Train a NeRF

import torch
from fast_ngp.models.fast_nerf import FastNGP_NeRF
from fast_ngp.utils.dataset import SimpleNeRFDataset
from fast_ngp.utils.trainer import NeRFTrainer

dataset_path = 'path/to/data'  # directory containing transforms_train.json or tiny_nerf_data.npz
dataset = SimpleNeRFDataset(root_dir=dataset_path, split='train', img_wh=(128, 128))
print(f"Dataset ready with {len(dataset)} rays")

# inti model
model = FastNGP_NeRF(
    encoding_config={
        'n_levels': 16,
        'n_features_per_level': 2,
        'log2_hashmap_size': 19,
        'base_resolution': 16,
        'finest_resolution': 512
    },
    mlp_config={
        'n_hidden_layers': 2,
        'hidden_dim': 64
    }
)
print("✅ Model initialized")

# Train

device = 'cuda' if torch.cuda.is_available() else 'cpu'
trainer = NeRFTrainer(
    model=model,
    train_dataset=dataset,
    val_dataset=None,  # you can add a validation dataset if available
    batch_size=64,
    lr=1e-2,
    num_epochs=5,
    device=device
)

trainer.train()
print("🎉 Training complete!")

# Provide a camera pose as a 3x4 matrix (numpy or torch)
import numpy as np

camera_pose = np.eye(4)[:3, :4]  # identity pose for testing
H, W = 128, 128

# The render_image method should be implemented in FastNGP_NeRF
with torch.no_grad():
    image = model.render_image(camera_pose, H=H, W=W)
print("Rendered image shape:", image.shape)

Command Line Interface

# Train NeRF
fast-ngp-train --task nerf --data path/to/data --config configs/nerf.yaml

# Render from trained model
fast-ngp-render --checkpoint model.pth --output renders/

Architecture

pyinstant-ngp/
├── encoding/          # Multiresolution hash encoding
├── models/          # MLP networks
├── rendering/         # Ray marching and rendering
├── example/              # Dataset loaders
└── utils/             # Dataset loaders and trainer

Citation

If you use this code in your research, please cite both the original paper and this implementation:

@article{mueller2022instant,
    title={Instant Neural Graphics Primitives with a Multiresolution Hash Encoding},
    author={M\"uller, Thomas and Evans, Alex and Schied, Christoph and Keller, Alexander},
    journal={ACM Transactions on Graphics (ToG)},
    volume={41},
    number={4},
    pages={1--15},
    year={2022},
    publisher={ACM}
}

@software{fast_ngp,
    author = {Chima Emmanuel},
    title = {Fast-NGP: Pure PyTorch Implementation of Instant-NGP},
    year = {2025},
    url = {https://github.com/Dox45/fast-ngp}
}

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgments

• Original Instant-NGP by NVIDIA
• Inspired by torch-ngp and HashNeRF-pytorch

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

Support

• 📖 Documentation
• 🐛 Issue Tracker
• 💬 Discussions