DragGaussian

Here we hardcode to use the hotdog scene in the following scripts.

1. Train 3DGS

Train 3DGS of the original scene and save as ply file.

python train.py \
    -s data/nerf_synthetic/hotdog \
    -m output/hotdog

2. Generate drag points and masks

python scripts/generate_handles.py

3. Train DragGaussian

python drag.py \
    -s data/nerf_synthetic/hotdog \
    -m output/hotdog \
    --start_ply_path output/hotdog/point_cloud/iteration_30000/point_cloud.ply \
    ---num_drag_steps 50 \
    --iterations 5000 \
    --eval \
    --output_path output/hotdog

3DGS

Installation

conda env create --file environment.yml
conda activate drag_gaussian

Dataset

Donwload nerf_synthetic dataset from this link and unzip it to data/nerf_synthetic/ directory.

Training

python train.py \
    -s data/nerf_synthetic/hotdog \
    -m output/hotdog

Render Round-view Video

python render_trajectory.py \
    -m output/hotdog \
    --num_frames 1000 \
    --output_video