synema 0.1

View Synthesis for Cinema

Synema

Synema is a collection of neural radiance field (NeRF-style) view-synthesis utilities designed to work with Cinema image-based databases for in situ / in transit visualization workflows.

The goal is to take an existing Cinema dataset (images + camera parameters) and enable novel view synthesis (new viewpoints, smooth camera paths, view interpolation) to enhance interactive exploration without requiring additional simulation output.

---

Backend: JAX / Flax Ecosystem

Synema is implemented using the JAX ecosystem, not PyTorch or TensorFlow.

Core ML stack:

• JAX --- numerical backend (XLA-compiled array computing)
• Flax --- neural network module system
• Optax --- optimizers
• Jaxtyping --- type annotations for JAX arrays

Why JAX?

• High-performance XLA compilation
• Functional programming model well-suited for NeRF training
• Clean separation between model definition (Flax) and optimization (Optax)
• Excellent support for GPU and TPU acceleration

GPU Support

To use a GPU, install the correct JAX build for your CUDA version:

pip install --upgrade "jax[cuda12]"

Refer to the official JAX installation guide for CUDA-specific wheels: https://jax.readthedocs.io/en/latest/installation.html

If no GPU is available, Synema will run on CPU (training will be slower).

---

What's in this repository

• synema/ --- Python package implementing models, data handling, and rendering utilities\
• examples/ --- Example scripts demonstrating training and inference on Cinema datasets\
• data/ --- Example assets and/or helper data\
• pyproject.toml --- Modern Python packaging (PEP 621 compliant)\
• requirements-dev.txt --- Development tooling dependencies\
• license.md --- Repository license

---

Requirements

Typical environment:

• Python 3.9+
• CUDA-capable GPU (recommended for training)
• JAX ecosystem (jax, flax, optax)
• Standard scientific Python libraries

---

Installation

Core install

pip install -e .

---

Quick Start

The repository includes two example Cinema databases:

• dragon_elevation.cdb --- scalar image database\
• tangle_rgb.cdb --- RGB image database

These are used by the example scripts in the examples/ directory:

• cinema_scalar_image.py → uses dragon_elevation.cdb
• cinema_rgb_image.py → uses tangle_rgb.cdb

The database paths are already hardcoded inside the example scripts.
You do not need to pass the database path via the command line.

Run Scalar Image Example

python examples/cinema_scalar_image.py

Run RGB Image Example

python examples/cinema_rgb_image.py

These scripts will:

1. Load the corresponding Cinema database\
2. Initialize the JAX/Flax NeRF-style model\
3. Train or evaluate depending on script configuration\
4. Produce rendered outputs / visualizations

Modify the script configuration directly if you want to change training parameters, output paths, or rendering settings.

---

Cinema Data Expectations

Synema expects:

• Images from a Cinema database
• Camera pose and intrinsics metadata
• Parameter metadata for consistent rendering

Adapters in synema/ or examples/ map Cinema metadata to NeRF training inputs.

---

Project Structure

synema/
  synema/        # Models, datasets, rendering, utilities
  examples/      # Training and rendering examples
  data/          # Example data
  pyproject.toml # Packaging configuration

---

Contributing

Contributions welcome via issues and pull requests:

• New model variants
• Dataset adapters
• Performance improvements
• Documentation updates

---

Citation

@software{synema,
  title  = {Synema: Novel View Synthesis for Cinema in situ Visualization},
  author = {CinemaScience contributors},
  url    = {https://github.com/cinemascience/synema},
  year   = {2025}
}

---

---

AI Assistance Disclosure

Portions of this repository's documentation, packaging configuration, and test scaffolding were generated or refined with the assistance of AI tools (including large language models).

All generated content has been reviewed and curated by the project maintainers. Any errors, omissions, or inconsistencies remain the responsibility of the human authors.

---

License

See license.md.