Fast Converging 3D Gaussian Splatting for 1-Minute Reconstruction

Ziyu Zhang, Tianle Liu, Diantao Tu, Shuhan Shen $^\dagger$

Institute of Automation, Chinese Academy of Sciences; Wuhan University

🏆 SIGA 2025 3DGS Challenge First Place 🏆

---

Overview

We present a fast 3DGS reconstruction pipeline designed to converge within one minute, developed for the SIGGRAPH Asia 3DGS Fast Reconstruction Challenge. Our solution includes several key points:

• Reverse per-Gaussian parallel optimization and compact forward splatting based on Taming-GS and Speedy-Splat;
• Load-balanced tiling writing in function duplicateWithKeys();
• Initialization from feed-forward 3DGS model based on AnySplat;
• Multi-view consistency-guided densification and pruning strategy inspired by Fast-GS;
• Supervise the rendered depth by depth estimator Metric3D;
• Optional global transform on camera extrinsics.

In addation, We tried using Neural Gaussians to replace original 3DGS for fast convergence with few parameters. Switch to the use_scaffold_release branch for details.

TODO

• Release the 3dgs code
• Add Load-balanced tiling writing
• Release the neural gaussian code

Getting Started

1. Prerequisites

• Ubuntu 22.04
• one 4090 GPU
• multi-core CPU (recommend > 20)

2. Installation

# 1) Create and activate a Conda environment
conda create -n 3dv_gs python=3.10
conda activate 3dv_gs

# 2) Install xFormers built for your CUDA
# Replace the URL with the one matching your local CUDA version. We use CUDA 12.4 and torch 2.5.1
# xformers==0.0.29 will automatically pull torch==2.5.1 as a dependency.
pip install -U xformers==0.0.29 --index-url https://download.pytorch.org/whl/cu124

# 3) Install torchvision that matches your torch
# torch 2.5.1 pairs with torchvision 0.20.1. We recommend using a prebuilt wheel:
# https://download.pytorch.org/whl/torchvision/
wget <the-correct-torchvision-wheel-from-the-link-above>
pip install torchvision-xxxx.whl

# 4) Verify versions and pin them
# Run `pip list` to confirm the installed torch/torchvision versions,
# then put those exact versions into constraints.txt.

# 5) Install project requirements with pinned versions (no build isolation)
pip install -r requirements.txt -c constraints.txt --no-build-isolation

# 6) Install local extensions (no build isolation)
pip install submodules/diff-gaussian-rasterization \
            submodules/simple-knn \
            submodules/fused-ssim \
            submodules/lanczos-resampling \
            --no-build-isolation

Also download the following weights/configs:

• AnySplat weights and config to ./anySplat/ckpt
• VGGT weights and config to ./vggt
• Metric3D v2 weights to ./metric3D/weight

Quick Start

1. Training

Modify Dataset Path

Edit .vscode/full_train_and_eval.sh:

• Set BASE_DIR to the root directory containing all scenes.
• Set EXP_DIR to your experiment output directory.

Then run training script for all scenes:

chmod +x .vscode/full_train_and_eval.sh
.vscode/full_train_and_eval.sh

Per-scene quantitative results are aggregated in ${EXP_DIR}/metrics_train.json, which records the rendering metrics and training time for all scenes.

Directory Layout Example

  $EXP_DIR
    ├── <scene_1>
    ...
    ├── <scene_n>
    └── metrics_train.json
  ...
  $BASE_DIR
    └── <scene_1>
        ├── images
        ├── sparse/0
        └── train_test_split.json

The file structure of train_test_split.json:

{
    "train": [
        "000000.png",
        "000003.png",
        ...,
        "000199.png"
    ],
    "test": [
        "000001.png",
        "000002.png",
        ...,
        "000193.png"
    ]
}

Heads-up on hardware variability

• Rare VRAM out-of-bounds on RTX 4090: We observed a very small probability of a memory out-of-bounds issue on 4090 GPUs, which we believe is caused by excessive VRAM pressure. If this occurs, please re-run the affected scene—in our tests, the issue is non-deterministic and typically disappears on retry.
• CPU Core Amount and Ruselts: The training loop will stop after 1 minute or when the maximum number of iterations is reached. We observe the training time to reach the same iteration may be different on cpus with different core numbers. Please modify the $iterations according to your device, so that the overall training time is about one minute.

2. Evaluation

Edit .vscode/full_eval.sh:

• Set DATA_PATH to the root directory containing all scenes.
• Set EVALUATE_DIR to same experiment output directory as above.

Then run eval script for all scenes:

chmod +x .vscode/full_eval.sh
.vscode/full_eval.sh

Please refer to metrics.json for the final time metrics and PSNR metrics.

Acknowledgments

We thank the above excellent open source repositories for sharing and the support of the competition organizers.

Citation

@article{zhang2026fastconverging3dgaussian,
  title={Fast Converging 3D Gaussian Splatting for 1-Minute Reconstruction}, 
  author={Ziyu Zhang and Tianle Liu and Diantao Tu and Shuhan Shen},
  journal={arXiv preprint arXiv:2601.19489},
  year={2026},
}