m80hz/g3splat

G³Splat: Geometrically Consistent Generalizable Gaussian Splatting

Model Description

G³Splat is a pose-free, self-supervised framework for generalizable Gaussian splatting that achieves state-of-the-art performance in:

• 🎯 Geometry Reconstruction - Accurate depth and mesh reconstructions
• 📐 Relative Pose Estimation - No camera poses required at inference
• 🎨 Novel View Synthesis - High-quality image rendering from new viewpoints

Available Checkpoints

Model	Gaussian Type	Training Data	File
G³Splat-3DGS	3DGS	RealEstate10K	g3splat_mast3r_3dgs_align_orient_re10k.ckpt
G³Splat-2DGS	2DGS	RealEstate10K	g3splat_mast3r_2dgs_align_orient_re10k.ckpt

Quick Start

from huggingface_hub import hf_hub_download

# Download 3DGS model
ckpt_path = hf_hub_download(
    repo_id="m80hz/g3splat",
    filename="g3splat_mast3r_3dgs_align_orient_re10k.ckpt"
)

# Or download 2DGS model
ckpt_path_2dgs = hf_hub_download(
    repo_id="m80hz/g3splat",
    filename="g3splat_mast3r_2dgs_align_orient_re10k.ckpt"
)

Usage

# Clone the repository
git clone https://github.com/m80hz/g3splat
cd g3splat

# Run demo
python demo.py --checkpoint pretrained_weights/g3splat_mast3r_3dgs_align_orient_re10k.ckpt

See the GitHub repository for full installation and usage instructions.

Training Details

• Training Data: RealEstate10K
• Resolution: 256×256
• Backbones: MASt3R (ViT-Large) and VGGT
• Hardware: 24× A100 GPUs (6 nodes × 4 GPUs)
• Training Time: ~6 hours

Citation

@inproceedings{g3splat,
  title     = {G3Splat: Geometrically Consistent Generalizable Gaussian Splatting},
  author    = {Hosseinzadeh, Mehdi and Chng, Shin-Fang and Xu, Yi and Lucey, Simon and Reid, Ian and Garg, Ravi},
  booktitle = {arXiv:2512.17547},
  year      = {2025},
  url       = {https://arxiv.org/abs/2512.17547}
}

License

MIT License