Lite Any Stereo: Efficient Zero-Shot Stereo Matching

This work presents Lite Any Stereo. It is a super efficient stereo matching model with strong zero-shot generalization ability. It outperforms or match accuracy-oriented models that do not use foundational priors, while requiring less than 1% of their computational cost.

Demo

Several example stereo image pairs are provided in the /assets/ directory.

You can visualize zero-shot stereo matching results of Lite Any Stereo on real-world scenes by running:

python demo.py

You can also test the model on your own stereo image pairs by replacing the input images.

Checkpoint

Before running the demo, please download the pretrained checkpoints from google drive . Then place them in: ./checkpoints/

Benchmark Results

To reproduce the benchmark results reported in Table 3 and Table 4 of the paper, run:

sh evaluate.sh

The results of Lite-CREStereo++ can be reproduced here.

MACs

To compute the model complexity (MACs), use:

python flops_count.py

Citation

If you find this work useful, please consider citing:

@article{jing2025lite,
  title={Lite Any Stereo: Efficient Zero-Shot Stereo Matching},
  author={Jing, Junpeng and Luo, Weixun and Mao, Ye and Mikolajczyk, Krystian},
  journal={arXiv preprint arXiv:2511.16555},
  year={2025}
}