Glob3R

Global Structure-from-Motion with 3D Foundation Models

Paper | GitHub | Project Page

Overview

This repository hosts the Glob3R checkpoint for geometry prediction and image matching.

Glob3R combines geometric foundation models with global optimization to reconstruct scenes from multiple images. It extends a frozen Pi3X backbone with learned dense matching, constructs multi-view tracks, and refines camera poses and scene geometry through motion averaging and bundle adjustment.

• Flexible inputs: supports ordered image sequences and unordered image collections.
• Scalable reconstruction: connects overlapping windows using keyframes and multi-view tracks.
• Global refinement: jointly optimizes geometric predictions to reduce drift and improve consistency.

Checkpoint

The file pytorch_model.bin contains the Glob3R geometry and matching weights. Use it with the Glob3R implementation; upstream Pi3/Pi3X checkpoints do not contain the matching heads trained for Glob3R.

Quick Start

1. Set up the code

Clone the repository and complete the installation guide, including its CUDA and optimization dependencies:

git clone --recurse-submodules https://github.com/aigc3d/Glob3R.git
cd Glob3R

2. Download the weights

From the Glob3R directory, run:

pip install -U huggingface_hub
hf download juneyoung9/glob3r pytorch_model.bin --local-dir weights

Also download the auxiliary SALAD and DINOv2 checkpoints listed in the weights section and place them in weights/.

3. Run reconstruction

Place an ordered image sequence in data/scene/, using filenames that sort into sequence order:

CUDA_VISIBLE_DEVICES=0 python infer.py \
  --data_path data/scene --save_path outputs/scene \
  --ckpt weights/pytorch_model.bin --weights_dir weights \
  --window_size 10 --anchor_imgs_per_window 1 \
  --keyframe_mode fixed --running_keyframe --save_dense

This configuration requires at least 11 selected images; reduce --window_size for shorter sequences. For unordered collections, add --unordered and remove --keyframe_mode fixed.

Further Information

• Training: matching-head fine-tuning using the released checkpoint.
• Evaluation: benchmark configurations and evaluation commands.
• Paper: method details and reported results.
• Issues: questions and bug reports.

Citation

If you use Glob3R in your research, please cite:

@misc{deng2026glob3r,
  title={Glob3R: Global Structure-from-Motion with 3D Foundation Models},
  author={Junyuan Deng and Heng Li and Kejie Qiu and Lingteng Qiu and Rui Peng and Weichao Shen and Weihao Yuan and Siyu Zhu and Zilong Dong and Ping Tan},
  year={2026},
  eprint={2607.09225},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2607.09225}
}