NeRF_Signature

Source code of the paper "The NeRF Signature: Codebook-Aided Watermarking for Neural Radiance Fields".

Installation

The project has the same dependencies as torch-ngp.

Usage

1. Train Clean Models

First, use torch-ngp to train clean NeRF models for different datasets:

Mip-NeRF 360

python main_nerf.py data/360_v2/counter --workspace logs/counter_clean -O --scale 0.33 --dt_gamma 0 --tcnn

Blender

python main_nerf.py data/nerf_synthetic/hotdog --workspace logs/hotdog_clean -O --bound 1.0 --scale 0.8 --dt_gamma 0 --tcnn

LLFF

python main_nerf.py data/nerf_llff_data/fern --workspace logs/fern_clean -O --tcnn

Tanks and Temple

python main_nerf.py data/TanksAndTemple/Family --workspace logs/Family_clean -O --bound 1.0 --scale 0.33 --dt_gamma 0 --tcnn

After training, move the .pth model files to the ./clean_model directory.

2. Train Watermarked Models

Train the signature representation:

Mip-NeRF 360

python main_nerf_wtmk.py data/360_v2/counter --workspace logs/git_counter_wtmk_32b -O --wtmk_tcnn --ckpt ./clean_model/counter_ngp_ep0125.pth --message_dim 32 --loss_w bce --lambda_w 0.005 --lambda_i 1.0 --num_rays 4096 --rand_pose 0 --n_views 1 --iters 1000 --num_rows 32 --num_cols 32 --use_existset --eval_interval 5 --save_interval 5 --num_images_test 10 --scale 0.33 --dt_gamma 0

Blender

python main_nerf_wtmk.py data/nerf_synthetic/hotdog --workspace logs/hotdog_wtmk_32b -O --wtmk_tcnn --ckpt ./clean_model/hotdog_ngp_ep0300.pth --message_dim 32 --downscale 2 --loss_w bce --lambda_w 0.005 --lambda_i 1.0 --num_rays 4096 --rand_pose 0 --n_views 1 --iters 1000 --num_rows 32 --num_cols 32 --use_existset --eval_interval 5 --save_interval 5 --num_images_test 10 --bound 1.0 --scale 0.8 --dt_gamma 0

LLFF

python main_nerf_wtmk.py data/nerf_llff_data/fern --workspace logs/fern_wtmk_32b -O --wtmk_tcnn --ckpt ./clean_model/fern_ngp_ep1500.pth --message_dim 32 --loss_w bce --lambda_w 0.005 --lambda_i 1.0 --num_rays 4096 --rand_pose 0 --n_views 1 --iters 600 --num_rows 32 --num_cols 32 --use_existset --eval_interval 10 --save_interval 10 --num_images_test 10

Tanks and Temple

python main_nerf_wtmk.py data/TanksAndTemple/Family --workspace logs/Family_wtmk_32b -O --wtmk_tcnn --ckpt ./clean_model/Family_ngp_ep0226.pth --message_dim 32 --loss_w bce --lambda_w 0.005 --lambda_i 1.0 --num_rays 4096 --rand_pose 0 --n_views 1 --iters 600 --num_rows 32 --num_cols 32 --use_existset --eval_interval 5 --save_interval 5 --num_images_test 10 --bound 1.0 --scale 0.33 --dt_gamma 0 --downscale 4

This project is built upon torch-ngp and HashNeRF-pytorch. We express our sincere gratitude to the authors of these repositories.

If you find our paper useful for your work please cite:

@article{luo2025signature,
  author    = {Ziyuan Luo and Anderson Rocha and Boxin Shi and Qing Guo and Haoliang Li and Renjie Wan},
  title     = {The NeRF Signature: Codebook-Aided Watermarking for Neural Radiance Fields},
  journal   = {IEEE Transactions on Pattern Analysis and Machine Intelligence},
  year      = {2025},
}