Xianqi Wang, Hao Yang, Hangtian Wang, Junda Cheng, Gangwei Xu, Min Lin, Xin Yang
Huazhong University of Science and Technology, Optics Valley Laboratory
- 05/14/2026: Update demo and training code.
- 04/15/2026: Update more versions of PromptStereo.
- 04/01/2026: Update the evaluation code.
conda create -n promptstereo python=3.12
conda activate promptstereo
pip install tqdm numpy wandb opt_einsum hydra-core
pip install imageio scipy torch torchvision opencv-python matplotlib
pip install xformers accelerate scikit-image
Data for evaluation:
Data for training:
| Model | Link |
|---|---|
| Depth-Anything-V2-Large | Download 🤗 |
| PromptStereo-SceneFlow-192 | Download 🤗 |
| PromptStereo-Unlimited-192 | Download 🤗 |
| PromptStereo-SceneFlow-576 | Download 🤗 |
| PromptStereo-Unlimited-576 | Download 🤗 |
The SceneFlow checkpoint is retrained and obtains better performance than the paper's version.
accelerate launch save_disparity.py
accelerate launch evaluate_stereo.py
# To evaluate with large disparity
accelerate launch evaluate_stereo.py checkpoint=checkpoint/unlimited_576.safetensors model.instance.cfg.gwc_max_disp=576
Default settings use bf16 precision with faster speed but a very little performance degration, you can set accelerator.mixed_precision to null to obtain entire performance.
# SceneFlow training
accelerate launch train_stereo.py
# Unlimited training
accelerate launch train_stereo.py --config-name train_unlimited checkpoint=checkpoint/sceneflow_192.safetensors
You should set tracker.init_kwargs.wandb.entity and save_path.
This project is based on Depth Anything V2, and MonSter. We thank the original authors for their excellent works.