Official implementation of Pow3R: Empowering Unconstrained 3D Reconstruction with Camera and Scene Priors (CVPR 2025)
[Project page], [arxiv]
@inproceedings{pow3r_cvpr25,
title={Pow3R: Empowering Unconstrained 3D Reconstruction with Camera and Scene Priors},
author={Wonbong Jang and Philippe Weinzaepfel and Vincent Leroy and Lourdes Agapito and Jerome Revaud},
booktitle = {CVPR},
year = {2025}
}see LICENSE
micromamba create -n pow3r python=3.11 cmake=3.14.0
micromamba activate pow3r
pip3 install torch==2.7.0 torchvision==0.22.0 --index-url https://download.pytorch.org/whl/cu126 # use the correct version of cuda for your system
pip install -r dust3r/requirements.txt
pip install -r dust3r/requirements_optional.txt
pip install -r requirements.txt
cd dust3r/croco/models/curope/
pip install . # or python setup.py build_ext --inplace
cd ../../../../We provide the pre-trained model:
| Modelname | Training resolutions | Head | Encoder | Decoder |
|---|---|---|---|---|
Pow3R_ViTLarge_BaseDecoder_512_linear.pth |
512x384, 512x336, 512x288, 512x256, 512x160 | Linear | ViT-L | ViT-B |
This is the same checkpoint as we used for CVPR.
We provide the code to generate the video showcasing the high-resolution capabilities of Pow3R.
python demo_high_res.py --img1 assets/img1.jpg --img2 assets/img2.jpg - TODO: release training code