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

This branch contains the code for the high-resolution (960x540) undistorted DL3DV dataset.

# create conda environment
conda create -n ilrm python=3.11 -y
conda activate ilrm

# install PyTorch (adjust cuda version according to your system)
pip install -r requirements.txt

The model checkpoints are host on HuggingFace.

This checkpoint was trained with 16//32 input images. We recommend finetuning it when using a different number of input images.

For training and evaluation, we used the DL3DV dataset after applying undistortion preprocessing with this script, originally introduced in Long-LRM.

Download the DL3DV benchmark dataset from here, and apply undistortion preprocessing.

Update the inference.ckpt_path field in configs/ilrm.yaml with the pretrained model.

Update the entries in data/dl3dv_eval.txt to point to the correct processed dataset path.

You can save videos or images by changing the fields (inference.save_video or save_images) in configs/ilrm.yaml.

The number of finetuning (post-prediction optimization) iterations is set to 10 in inference.finetune_iter.

# inference
CUDA_VISIBLE_DEVICES=0 python inference.py --config configs/ilrm.yaml

# post-prediction optimization
CUDA_VISIBLE_DEVICES=0 python finetine.py --config configs/ilrm.yaml

@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}
}

This branch is built on many amazing research works, thanks a lot to all the authors for sharing!

• Gaussian-Splatting and gsplat
• LVSM
• Long-LRM
• LaCT