TSDN: Transport-based Stylization for Dynamic NeRF

This repository contains a pipeline for dynamic scene stylization using TSDN based on TiNeuVox. The dataset is converted from Robust Dynamic Radiance Fields to a D-NeRF compatible format.

Key Limitations

Our current implementation faces constraints when processing:

• Long-duration videos
• Scenes with drastic motions

These limitations stem from fundamental challenges in Dynamic NeRF architectures. We explicitly encourage researchers to explore solutions in these directions and welcome contributions to overcome these constraints.

Quick Start

1. Dynamic Reconstruction
   Reconstruct the dynamic NeRF scene:

   python run.py

2. Stylization Generation
   Obtain stylized results using:

   python run_TSDN.py

Implementation Notes

⚠️ Current Constraints

Core parameters are hardcoded in run.py and run_TSDN.py (to be modularized in future releases)

Dataset conversion tools will be released after code refactoring

Requirements

• lpips
• mmcv
• imageio
• imageio-ffmpeg
• opencv-python
• pytorch_msssim
• torch
• torch_scatter