Lite3R: A Model-Agnostic Framework for Efficient Feed-Forward 3D Reconstruction

Official implementation of Lite3R, a model-agnostic framework for efficient feed-forward 3D reconstruction from multi-view images.

Lite3R: A Model-Agnostic Framework for Efficient Feed-Forward 3D Reconstruction

Haoyu Zhang*, Zeyu Zhang*^†, Zedong Zhou, Yang Zhao, and Hao Tang^#

*Equal contribution. ^†Project lead. ^#Corresponding author.

Paper | Website | Models | HF Paper

Overview

Lite3R introduces a systematic approach to compress large-scale 3D reconstruction models while maintaining reconstruction quality. The framework combines:

• Sparse Linear Attention (SLA): Efficient attention mechanism that reduces computational complexity
• FP8-Aware Quantization-Aware Training (QAT): Low-precision training for deployment efficiency
• Partial Attention Distillation: Knowledge transfer from dense teacher models

The framework has been validated on two state-of-the-art architectures:

• VGGT (Visual Geometry Grounding Transformer)
• Depth Anything V3 Large (DA3-L)

Installation

# Clone the repository
git clone https://github.com/AIGeeksGroup/Lite3R.git
cd Lite3R

# Create conda environment
conda create -n lite3r python=3.10
conda activate lite3r

# Install dependencies
pip install -r requirements.txt

Model Checkpoints

Pre-trained model weights are available on Hugging Face:

• vggt_fp8_qat_1ep.pt - Lite3R VGGT
• da3_fp8_qat_1ep.pt - Lite3R DA3-L

Download and place checkpoints in checkpoints/fp8_qat_1ep/.

Quick Start

Inference

python inference.py \
  --model vggt \
  --checkpoint checkpoints/fp8_qat_1ep/vggt/vggt_fp8_qat_1ep.pt \
  --input_dir examples/input \
  --output examples/output/reconstruction.ply

Training

VGGT backbone:

python train/train_vggt.py --config configs/final/vggt_fp8_qat_1ep.yaml

DA3-L backbone:

python train/train_da3.py --config configs/final/da3_fp8_qat_1ep.yaml

Evaluation

Evaluate on BlendedMVS or DTU datasets:

python eval/eval_vggt.py \
  --config configs/final/vggt_eval_blended.yaml \
  --ckpt checkpoints/fp8_qat_1ep/vggt/vggt_fp8_qat_1ep.pt \
  --name vggt_lite3r

Citation

If you find this work useful, please cite:

@article{zhang2026lite3r,
  title={Lite3R: A Model-Agnostic Framework for Efficient Feed-Forward 3D Reconstruction},
  author={Zhang, Haoyu and Zhang, Zeyu and Zhou, Zedong and Zhao, Yang and Tang, Hao},
  journal={arXiv preprint arXiv:2605.11354},
  year={2026}
}

License

This project is released under the MIT License. See LICENSE for details.

Acknowledgments

This work builds upon VGGT and Depth Anything V3.