EDGS: Eliminating Densification for Efficient Convergence of 3DGS

Dmytro Kotovenko^* · Olga Grebenkova^* · Björn Ommer

CompVis @ LMU Munich · Munich Center for Machine Learning (MCML)

* equal contribution

3DGS initializes with a sparse set of Gaussians and progressively adds more in under-reconstructed regions. In contrast, EDGS starts with a dense initialization from triangulated 2D correspondences across training image pairs, requiring only minimal refinement. This leads to faster convergence and higher rendering quality. Our method reaches the original 3DGS LPIPS score in just 25% of the training time and uses only 60% of the splats. Renderings become nearly indistinguishable from ground truth after only 3,000 steps — without any densification.

3D scene reconstruction using our method in 11 seconds.

📚 Table of Contents

• 🚀 Quickstart
• 🛠️ Installation

• 📦 Data

• 🏋️ Training

• 🏗️ Reusing Our Model

• 📄 Citation

🚀 Quickstart

The fastest way to try our model is through the Hugging Face demo, which lets you upload images or a video and interactively rotate the resulting 3D scene. For broad accessibility, we currently support only forward-facing scenes.

Steps:

1. Upload a list of photos or a single video.
2. Click 📸 Preprocess Input to estimate 3D positions using COLMAP.
3. Click 🚀 Start Reconstruction to run the model.

You can also explore the reconstructed scene in 3D directly in the browser.

⚡ Runtime: EDGS typically takes just 10–20 seconds, plus 5–10 seconds for COLMAP processing. Additional time may be needed to save outputs (model, video, 3D preview).

You can also run the same app locally on your machine with command: CUDA_VISIBLE_DEVICES=0 python gradio_demo.py --port 7862 --no_share Without --no_share flag you will get the adress for gradio app that you can share with the others allowing others to process their data on your server.

Alternatively, check our Colab notebook.

🛠️ Installation

You can either run install.sh or manually install using the following:

```bash git clone git@github.com:CompVis/EDGS.git --recursive cd EDGS git submodule update --init --recursive

conda create -y -n edgs python=3.10 pip conda activate edgs

Set up path to your CUDA. In our experience similar versions like 12.2 also work well

export CUDAHOME=/usr/local/cuda-12.1 export LDLIBRARYPATH=$CUDAHOME/lib64:$LDLIBRARYPATH export PATH=$CUDA_HOME/bin:$PATH

conda install pytorch torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidia -y conda install nvidia/label/cuda-12.1.0::cuda-toolkit -y

pip install -e submodules/gaussian-splatting/submodules/diff-gaussian-rasterization pip install -e submodules/gaussian-splatting/submodules/simple-knn

For COLMAP and pycolmap

Optionally install original colmap but probably pycolmap suffices

conda install conda-forge/label/colmap_dev::colmap

pip install pycolmap

pip install wandb hydra-core tqdm torchmetrics lpips matplotlib rich plyfile imageio imageio-ffmpeg conda install numpy=1.26.4 -y -c conda-forge --override-channels

pip install -e submodules/RoMa conda install anaconda::jupyter --yes

Stuff necessary for gradio and visualizations

pip install gradio pip install plotly scikit-learn moviepy==2.1.1 ffmpeg pip install open3d ```

📦 Data

We evaluated on the following datasets:

• MipNeRF360 — download here. Unzip "Dataset Pt. 1" and "Dataset Pt. 2", then merge scenes.
• Tanks & Temples + Deep Blending — from the original 3DGS repo.

Using Your Own Dataset

You can use the same data format as the 3DGS project. Please follow their guide to prepare your scene.

Expected folder structure: scene_folder |---images | |---<image 0> | |---<image 1> | |---... |---sparse |---0 |---cameras.bin |---images.bin |---points3D.bin

Nerf synthetic format is also acceptable.

You can also use functions provided in our code to convert a collection of images or a sinlge video into a desired format. However, this may requre tweaking and processing time can be large for large collection of images with little overlap.

🏋️ Training

To optimize on a single scene in COLMAP format use this code.
bash python train.py \ train.gs_epochs=30000 \ train.no_densify=True \ gs.dataset.source_path=<scene folder> \ gs.dataset.model_path=<output folder> \ init_wC.matches_per_ref=20000 \ init_wC.nns_per_ref=3 \ init_wC.num_refs=180

Command Line Arguments for train.py

• train.gs_epochs Number of training iterations (steps) for Gaussian Splatting.
• train.no_densify Disables densification. True by default.
• gs.dataset.source_path Path to your input dataset directory. This should follow the same format as the original 3DGS dataset structure.
• gs.dataset.model_path Output directory where the trained model, logs, and renderings will be saved.
• init_wC.matches_per_ref Number of 2D feature correspondences to extract per reference view for initialization. More matches leads to more gaussians.
• init_wC.nns_per_ref Number of nearest neighbor images used per reference during matching.
• init_wC.num_refs Total number of reference views sampled.

• wandb.mode Specifies how Weights & Biases (W&B) logging is handled.

  • Default: "disabled"
  • Options:
    • "online" — log to the W&B server in real-time
    • "offline" — save logs locally to sync later
    • "disabled" — turn off W&B logging entirely

  If you want to enable W&B logging, make sure to also configure: - wandb.project — the name of your W&B project - wandb.entity — your W&B username or team name

Example override: bash wandb.mode=online wandb.project=EDGS wandb.entity=your_username train.gs_epochs=15_000 init_wC.matches_per_ref=15_000

To run full evaluation on all datasets:

bash python full_eval.py -m360 <mipnerf360 folder> -tat <tanks and temples folder> -db <deep blending folder>

🏗️ Reusing Our Model

Our model is essentially a better initialization module for Gaussian Splatting. You can integrate it into your pipeline by calling:

python source.corr_init.init_gaussians_with_corr(...)

Input arguments:

• A GaussianModel and Scene instance
• A configuration namespace cfg.init_wC to specify parameters like the number of matches, neighbors, and reference views
• A RoMA model (automatically instantiated if not provided)

📄 Citation

```bibtex @misc{kotovenko2025edgseliminatingdensificationefficient, title={EDGS: Eliminating Densification for Efficient Convergence of 3DGS}, author={Dmytro Kotovenko and Olga Grebenkova and Björn Ommer}, year={2025}, eprint={2504.13204}, archivePrefix={arXiv}, primaryClass={cs.GR}, url={https://arxiv.org/abs/2504.13204}, }

```

TODO:

• [ ] Code for training and processing forward-facing scenes.
• [ ] More data examples