Lite Any Stereo Series

Official codebase for the Lite Any Stereo (LAS) series.

This repository supports LAS1 and LAS2 S/M/L/H release models.

Overview

Lite Any Stereo is a series of efficient zero-shot stereo matching models for practical deployment. This repository contains the public evaluation and inference code for LAS1 and LAS2.

Performance Snapshot

Zero-shot performance and runtime comparison. Runtime is reported on H200 / Orin 8G.

Checkpoints

The pretrained checkpoints are hosted on Hugging Face. A Google Drive mirror is also available here. Please place the downloaded checkpoints in ./checkpoints/. The release uses the following default filenames:

LAS2 defaults to the M model when --model_size is not specified. You can always pass a checkpoint explicitly with --restore_ckpt.

Demo

Several side-by-side stereo image pairs are provided in ./assets/. Pass --stereo_file to use another pair. Run LAS1:

python demo.py --version las1 --restore_ckpt ./checkpoints/LiteAnyStereo.pth

Run LAS2-M:

python demo.py --version las2 --model_size m --restore_ckpt ./checkpoints/LAS2_M.pth

Run another LAS2 release model by changing --model_size:

python demo.py --version las2 --model_size h --restore_ckpt ./checkpoints/LAS2_H.pth

The demo saves the disparity visualization, raw disparity array, and optional point-cloud outputs to --out_dir.

Evaluation

To reproduce the benchmark evaluation commands, run:

VERSION=las1 sh evaluate.sh
VERSION=las2 MODEL_SIZE=s sh evaluate.sh
VERSION=las2 MODEL_SIZE=m sh evaluate.sh
VERSION=las2 MODEL_SIZE=l sh evaluate.sh
VERSION=las2 MODEL_SIZE=h sh evaluate.sh

You can also evaluate one dataset directly:

python evaluate_stereo.py --version las2 --model_size h --restore_ckpt ./checkpoints/LAS2_H.pth --dataset middlebury_H

Supported datasets are middlebury_F, middlebury_H, middlebury_Q, eth3d, kitti, and drivingstereo.

MACs

To compute model complexity:

python flops_count.py --version las1
python flops_count.py --version las2 --model_size m
python flops_count.py --version las2 --model_size h

Runtime

To measure inference time:

python profile_speed.py --version las1
python profile_speed.py --version las2 --model_size m
python profile_speed.py --version las2 --model_size h

The runtime script uses CUDA synchronization when running on GPU.

Citation

If you find the released code useful, please consider citing:

@InProceedings{jing2026litestereo,
    author    = {Jing, Junpeng and Luo, Weixun and Mao, Ye and Mikolajczyk, Krystian},
    title     = {Lite Any Stereo: Efficient Zero-Shot Stereo Matching},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2026},
    pages     = {21725-21735}
}

@article{jing2026litestereov2,
      title={Lite Any Stereo V2: Faster and Stronger Efficient Zero-Shot Stereo Matching}, 
      author={Junpeng Jing and Ronglai Zuo and Zhelun Shen and Shangchen Zhou and Rolandos Alexandros Potamias and Stefanos Zafeiriou and Krystian Mikolajczyk and Jiankang Deng},
      year={2026},
      eprint={2606.24457},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2606.24457}, 
}