ShinyNeRF

ShinyNeRF models anisotropic reflections in Neural Radiance Fields, capturing view-dependent effects whose shape/orientation changes with surface rotation (e.g., brushed metals, satin/velvet, nacre).

• Built on top of NeRF-Factory (you can still run other methods via their .gin configs).
• Primary focus: anisotropic material modeling in NeRFs.

Paper: https://arxiv.org/abs/2512.21692
Project page: https://multimedia-eurecat.github.io/ShinyNeRF/

The project was implemented by Albert Barreiro.

Installation

conda create -n shinynerf python=3.8.20 -y
conda activate shinynerf

# install torch + cuda (Conda)
conda install -c pytorch -c conda-forge \
  pytorch==1.11.0 torchvision==0.12.0 torchaudio==0.11.0 cudatoolkit==11.3.1 -y

# IMPORTANT: pin pip (some deps have metadata incompatibilities with newer pip)
python -m pip install "pip<24.1"

# install repo requirements
pip install -r requirements.txt

Data format

Each scene directory must contain:

• train/ images
• test/ images
• transforms_train.json
• transforms_test.json

Recommended layout (matches --scene teapot):

data/
  teapot/
    train/
      000.png
      001.png
      ...
    test/
      000.png
      001.png
      ...
    transforms_train.json
    transforms_test.json

Training

python3 run.py --ginc configs/[model]/[data].gin --scene [scene] --downscale [value]

# Example:

python3 run.py \
  --ginc configs/shinynerf/shiny_blender_bitangent_radius.gin \
  --scene teapot \
  --downscale 1

Where:

• --scene teapot expects data/teapot/ (see Data format).
• --downscale controls resolution scaling (1 = full res, 2 = half, 4 = quarter, ...).

Evaluation

Evaluation does not support multi-GPU (run it on a single GPU / single process).

Also note that --eval will save a lot of auxiliary outputs (not just the final RGB), e.g. RGB, depth, normals, tangents/bitangents, etc. This can quickly take significant disk space, especially for high resolutions / many views.

python3 run.py --ginc configs/[model]/[data].gin --scene [scene] --downscale [value] --eval

# Example:

python3 run.py \
  --ginc configs/shinynerf/shiny_blender_bitangent_radius.gin \
  --scene teapot \
  --downscale 1 \
  --eval

Outputs

Runs will create an experiment log directory (exact naming depends on the NeRF-Factory logging setup + config). To monitor training:

tensorboard --logdir logs --samples_per_plugin images=100

Troubleshooting

pip / dependency metadata errors

Make sure pip is:

python -m pip install "pip<24.1"

License

This repository includes original code for ShinyNeRF:

Copyright (c) 2026 Albert Barreiro

and builds on top of NeRF-Factory, whose code and associated portions remain:

Copyright (c) 2022 POSTECH, KAIST, and Kakao Brain Corp.

Licensed under the Apache License, Version 2.0 (Apache-2.0).
See LICENSE for the full license text. If a NOTICE file is present, it must be preserved when redistributing.

Citation

@Article{Barreiro2025ShinyNeRF,
  author    = {Barreiro, Albert and Mar\'{i}, Roger and Redondo, Rafael and Haro, Gloria and Bosch, Carles},
  title     = {{ShinyNeRF}: Modeling Anisotropic Reflections in Neural Radiance Fields},
  journal   = {The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences},
  year      = {2026}
}

Acknowledgements

This project is built on top of NeRF-Factory and uses its training/evaluation pipeline and configuration system.