Home

Awesome

Seal-3D: Interactive Pixel-Level Editing for Neural Radiance Fields

teaser

The official implementation of the paper Seal-3D: Interactive Pixel-Level Editing for Neural Radiance Fields, the first interactive pixel-level NeRF editing tool.

Accepted by ICCV 2023.

Project PagePaperArXivCode

This project is built on ashawkey/torch-ngp's NGP and TensoRF implementation.

Installation

To find more details about the development environment setup, please refer to torch-ngp#install.

git clone --recursive https://github.com/windingwind/seal-3d.git
cd seal-3d

Install with pip

pip install -r requirements.txt

# (optional) install the tcnn backbone
pip install git+https://github.com/NVlabs/tiny-cuda-nn/#subdirectory=bindings/torch

Install with conda

conda env create -f environment.yml
conda activate torch-ngp

Build extension (optional)

By default, we use load to build the extension at runtime. However, this may be inconvenient sometimes. Therefore, we also provide the setup.py to build each extension:

# install all extension modules
bash scripts/install_ext.sh

# if you want to install manually, here is an example:
cd raymarching
python setup.py build_ext --inplace # build ext only, do not install (only can be used in the parent directory)
pip install . # install to python path (you still need the raymarching/ folder, since this only install the built extension.)

Dataset

We use the same data format as instant-ngp. Please download and put them under ./data.

To find more details about the supported dataset, please refer to torch-ngp#usage.

Usage

Code Structure

Based on the implementation of the repo, we slightly modified the files in nerf (the NGP implementation) and tensoRF (the TensoRF implementation) to fit our needs.

The main entrances are main_SealNeRF.py (NGP backbone) and main_SealTensoRF.py (TensoRF backbone).

In SealNeRF:

Train

Follow the steps below to apply the editing operation on an existing NeRF model:

  1. Train an NGP/TensoRF model following the instructions of torch-ngp#usage. For example:
# NGP backbone, Lego
python main_nerf.py data/nerf_synthetic/lego/ --workspace exps/lego_ngp -O --bound 1.0 --scale 0.8 --dt_gamma 0
  1. Train Seal3D on the model you get in the previous step (headless mode).
# Headless mode, bounding shape editing, NGP backbone, Lego
# pretraining_epochs: pretraining stage epochs
# extra_epochs: total epochs (pretraining + finetuning)
# pretraining_*_point_step: pretraining sample step
# ckpt: the input student model checkpoint
# teacher_workspace: teacher model workspace
# teacher_ckpt: teacher model checkpoint
# seal_config: the editing config directory used in headless mode. the config file is $seal_config/seal.json.
# eval_interval & eval_count: control eval behavior
python main_SealNeRF.py data/nerf_synthetic/lego/\
    --workspace exps/lego_ngp_bbox -O --bound 1.0 --scale 0.8 --dt_gamma 0\
    --pretraining_epochs 100 --extra_epochs 150\
    --pretraining_local_point_step 0.005 --pretraining_surrounding_point_step -1\
    --pretraining_lr 0.05 --ckpt exps/lego_ngp/checkpoints/ngp_ep0300.pth\
    --teacher_workspace exps/lego_ngp --teacher_ckpt exps/lego_ngp/checkpoints/ngp_ep0300.pth\
    --seal_config data/seal/lego_bbox/\
    --eval_interval 100 --eval_count 10

The seal_config files used by examples in the paper can be downloaded from Google Drive link. The explanation for the parameters of seal_config can be found in the corresponding class of the proxy function in SealNeRF/seal_utils.

The full argument list and descriptions can be found in the corresponding entrance file (main_*.py).

To start in GUI mode, use --gui.

Currently, GUI mode supports Color, Anchor, Brush, and Texture editing.

BibTeX

@misc{wang2023seal3d,
      title={Seal-3D: Interactive Pixel-Level Editing for Neural Radiance Fields}, 
      author={Xiangyu Wang and Jingsen Zhu and Qi Ye and Yuchi Huo and Yunlong Ran and Zhihua Zhong and Jiming Chen},
      year={2023},
      eprint={2307.15131},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Acknowledgement

Use this code under the MIT License. No warranties are provided. Keep the laws of your locality in mind!

Please refer to torch-ngp#acknowledgement for the acknowledgment of the original repo.