ILRM

<div align="center"> <h1><span style="color: rgb(230, 100, 80);">i</span><span style="color: rgb(230, 183, 53);">L</span><span style="color: rgb(117, 160, 85);">R</span><span style="color: rgb(96, 120, 172);">M</span>: An Iterative Large 3D Reconstruction Model</h1>

<a href="https://arxiv.org/abs/2507.23277"><img src="https://img.shields.io/badge/arXiv-2507.23277-b31b1b" alt="arXiv"></a> <a href="https://gynjn.github.io/iLRM/"><img src="https://img.shields.io/badge/Project_Page-green" alt="Project Page"></a>

Gyeongjin Kang, Seungtae Nam, Xiangyu Sun, Sameh Khamis, Abdelrahman Mohamed, Eunbyung Park

</div>

Official repo for the paper "iLRM: An Iterative Large 3D Reconstruction Model"

Check out the wide branch for wide-coverage scene reconstruction on the DL3DV dataset!

Installation

# create conda environment
conda create -n ilrm python=3.10 -y
conda activate ilrm

# install PyTorch (adjust cuda version according to your system)
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt

Checkpoints

We first release the 2-view RealEstate10K (256x256) checkpoint, which is the most common baseline in related works and provides a standard reference point for comparison. We will upload other checkpoints soon!

In training and evaluation, we used the dataset preprocessed by pixelSplat.

The model checkpoints are host on HuggingFace.

| Model | PSNR | SSIM | LPIPS | | ----- | ----- | ----- | ----- | | re10k_2view | 28.65 | 0.900 | 0.110 |

Inference

Update the dataset.roots field in config/experiment/re10k.yaml with your dataset path.

Update the checkpointing.load field in config/main.yaml with the pretrained model.

CUDA_VISIBLE_DEVICES=0 python -m src.main +experiment=re10k dataset/view_sampler=evaluation dataset.view_sampler.index_path=assets/evaluation_index_re10k.json

Citation

@article{kang2025ilrm,
  title={iLRM: An Iterative Large 3D Reconstruction Model},
  author={Kang, Gyeongjin and Nam, Seungtae and Sun, Xiangyu and Khamis, Sameh and Mohamed, Abdelrahman and Park, Eunbyung},
  journal={arXiv preprint arXiv:2507.23277},
  year={2025}
}

ACKNOWLEDGEMENTS

This work was supported by Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (RS-2025-02653113, High-Performance Research AI Computing Infrastructure Support at the 2 PFLOPS Scale)