Flash3D

Production-quality 3D Vision: Gaussian Splatting, NeRF, Depth Estimation & 3D Reconstruction

---

Overview

Flash3D is a unified Python framework for state-of-the-art 3D vision, combining Gaussian Splatting, Neural Radiance Fields, monocular depth estimation, and 3D reconstruction into a single, modular library. Built for researchers and engineers who need production-ready 3D pipelines.

Key Features

• 3D Gaussian Splatting – Per-scene optimization with adaptive density control
• Feed-Forward 3DGS – Single-pass reconstruction without per-scene training (pixelSplat/YoNoSplat-style)
• Neural Radiance Fields – MLP-based volume rendering with hash encoding (instant-NGP)
• Monocular Depth – Foundation model-compatible depth estimation
• Multi-format Export – PLY, OBJ, ONNX, .splat for web viewers
• Comprehensive Metrics – PSNR, SSIM, LPIPS, Chamfer Distance, F1
• LoRA Fine-tuning – Parameter-efficient adaptation for domain transfer

---

Installation

# Standard
pip install -e .

# Full (with visualization, metrics, experiment tracking)
pip install -e ".[full]"

# Development
pip install -e ".[dev,full]"

Or use the setup script:

chmod +x setup_env.sh && ./setup_env.sh

---

Quick Start

Scene Reconstruction

from flash3d import SceneReconstructor

reconstructor = SceneReconstructor(method="gaussian_splatting")
result = reconstructor.reconstruct("path/to/colmap/scene/", num_iterations=30_000)

Depth Estimation

from flash3d import DepthEstimator

estimator = DepthEstimator()
estimator.predict("images/", "depth_output/")

Novel View Synthesis

from flash3d import ViewSynthesizer

synth = ViewSynthesizer.from_checkpoint("model.pth")
frames = synth.render_orbit(num_frames=120, output_dir="renders/")

CLI

flash3d train --config configs/flash3d_gaussian_splatting.yaml
flash3d render --checkpoint model.pth --num-frames 120
flash3d depth --input images/ --output depth/
flash3d export --checkpoint model.pth --format ply
flash3d benchmark --checkpoint model.pth --dataset data/test/

---

Supported Models

Model	Type	Use Case	Paper
3D Gaussian Splatting	Per-scene	High-quality novel views	Kerbl et al. 2023
NeRF (hash encoding)	Per-scene	Volume rendering	Müller et al. 2022
Feed-Forward 3DGS	Generalizable	Single-pass reconstruction	Charatan et al. 2024
Monocular Depth	Single-image	Depth from RGB	Yang et al. 2024

---

Project Structure

Flash3D/
├── flash3d/               # Core library
│   ├── models/            # Model architectures (3DGS, NeRF, FF-3DGS)
│   ├── rendering/         # Differentiable rasterization & ray marching
│   ├── geometry/          # Point clouds, depth, meshes, transforms
│   ├── engine/            # Training, validation, prediction, export
│   ├── tasks/             # Task definitions (NVS, reconstruction, depth)
│   ├── solutions/         # High-level APIs
│   ├── analytics/         # Benchmarking & metrics
│   └── utils/             # I/O, visualization, callbacks
├── configs/               # YAML configuration files
├── examples/              # Usage examples
├── tests/                 # Pytest test suite
├── docs/                  # Documentation
└── docker/                # Containerization

---

Supported Datasets

Dataset	Format	Scenes	Use
COLMAP	Sparse reconstruction	Any	Per-scene training
ScanNet	RGBD sequences	1500+ indoor	Depth & reconstruction
RealEstate10K	Video frames	10K sequences	Wide-baseline NVS
DL3DV	Multi-view captures	10K diverse	Large-scale training

---

Configuration

Flash3D uses YAML-based configuration:

task: novel_view_synthesis
model:
  name: gaussian_splatting
  num_gaussians: 100000
  sh_degree: 3
data:
  dataset: colmap
  root_dir: data/garden/
train:
  max_iterations: 30000
  densify_grad_threshold: 0.0002

---

Benchmarks

Evaluated on Mip-NeRF 360 (outdoor scenes):

Method	PSNR ↑	SSIM ↑	LPIPS ↓	Train Time
3DGS	27.5	0.815	0.220	~25 min
NeRF (hash)	26.8	0.790	0.250	~15 min
FF-3DGS	24.2	0.750	0.300	~0.1s/scene

---

Development

# Run tests
pytest tests/ -v

# Lint
ruff check flash3d/

# Type check
mypy flash3d/

---

Citation

@software{flash3d2026,
  title={Flash3D: Production-quality 3D Vision Library},
  author={FlashVision},
  year={2026},
  url={https://github.com/FlashVision/FlashVision/tree/main/Flash3D}
}

---

License

This project is licensed under the MIT License.