Nuvo

Personal Implementation of the paper: Nuvo: Neural UV Mapping for Unruly 3D Representations

Dependencies

Tested environment:

• Linux Ubuntu 20.04
• RTX 4090

To install the required libraries, run:

conda create -n nuvo python=3.9
conda activate nuvo
pip install -r requirements.txt

I used Pytorch 2.3.0 with CUDA 11.8. Other versions may work, but are not tested.

Training

python train.py --config configs/<config-name.json>

Results

Below are the predicted UV mappings for the bunny mesh.

1. Bunny

bunny.mp4

Features

In the default Nuvo implementation, some triangles may have vertices that span different UV charts. This can be problematic for rendering and texture mapping, as there will be interpolation across the charts (see the first and third rows of the below figure). To address this issue, we implemented a vertex duplication feature that prevents this interpolation (see the second and fourth rows).

The vertex duplication is used by default. To disable it, set use_vertex_duplication to false in the config file.

Acknowledgement

This repository is based on the following paper / resources:

• Nuvo: Neural UV Mapping for Unruly 3D Representations

  @article{srinivasan2023nuvo,
    author    = {Pratul P. Srinivasan and Stephan J. Garbin and Dor Verbin and Jonathan T. Barron and Ben Mildenhall},
    title     = {Nuvo: Neural UV Mapping for Unruly 3D Representations},
    journal = {arXiv},
    year      = {2023},
  }

• Alec Jacobson's collection of 3D models