Glob3R: Global Structure-from-Motion With 3D Foundation Models

NeurIPS 2026

HKUST Spatial Artificial Intelligence Lab · Alibaba Group · Nanjing University · Fudan University

Junyuan Deng, Heng Li, Kejie Qiu, Lingteng Qiu, Rui Peng, Weichao Shen, Weihao Yuan, Siyu Zhu, Zilong Dong, Ping Tan

Overview

Glob3R is a scalable global 3D reconstruction framework built on geometric foundation models. It converts dense feed-forward predictions into reliable multi-view tracks and jointly refines camera poses and scene geometry through motion averaging and bundle adjustment. Glob3R improves reconstruction accuracy and consistency on long sequences, large-scale scenes, and unordered image collections.

Installation · Training · Evaluation

TODO

• Release inference code.
• Release evaluation code.
• Release training code.
• Improve depth alignment: aligning depth maps with scale and shift alone may be insufficient for consistent point cloud fusion.

Installation

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

Follow installation to create the CUDA environment and build the pinned optimization dependencies.

Weights

Place these files in weights/:

File	Purpose	Source
pytorch_model.bin	Glob3R geometry and matching	Hugging Face
dino_salad.ckpt	Image retrieval	SALAD
dinov2_vitb14_pretrain.pth	Retrieval backbone	DINOv2
dinov2_vitb14_reg4_pretrain.pth	Keyframe selection	DINOv2 with registers

Download the Glob3R checkpoint:

mkdir -p weights
curl -fL https://huggingface.co/juneyoung9/glob3r/resolve/main/pytorch_model.bin \
  -o weights/pytorch_model.bin

Use the Glob3R checkpoint; upstream Pi3/Pi3X weights do not contain its trained matching heads. The two DINOv2 checkpoints serve different architectures.

Inference

For an ordered sequence, put images in one directory with sortable names:

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

For an unordered collection, use --unordered and omit --keyframe_mode. Window size 10 processes up to 20 views per ordered window and requires at least 11 selected images. For a shorter sequence, reduce --window_size. Benchmark-specific settings and evaluation commands are in evaluation.

--ckpt and --weights_dir can also be set through GLOB3R_CKPT and GLOB3R_WEIGHTS. Run python infer.py --help for options.

Training

Fine-tune the matching heads using public TUM RGB-D sequences:

python -m glob3r.training.train --config configs/train_tum.yaml \
  --data-root data/tum --checkpoint weights/pytorch_model.bin \
  --output outputs/tum_refine

Training covers data layout, online warp supervision, validation and checkpoint recovery. This example continues from the released model; it does not recreate the original mixed-dataset training run.

Acknowledgements

Thanks to these great repositories:

Sail-Recon for the PSNR evaluation;

Pi3, RomaV2, Instant-SfM, Glomap for inspiration in algorithm design and implementation;

And many other inspiring works in the community.

Original contributions use the BSD 3-Clause license. Incorporated code retains its own terms, including GPL and noncommercial restrictions; the complete project is not uniformly BSD-licensed. See third-party attribution and licenses.

Citation

If you find this project useful in your research, please consider citing:

@inproceedings{glob3r2026,
  title={Glob3R: Global Structure-from-Motion with 3D Foundation Models},
  author={Deng, Junyuan and Li, Heng and Qiu, Kejie and Qiu, Lingteng and Peng, Rui and Shen, Weichao and Yuan, Weihao and Zhu, Siyu and Dong, Zilong and Tan, Ping},
  booktitle={Advances in Neural Information Processing Systems},
  year={2026}
}