FoundationStereo is a foundation model developed by NVIDIA Research for stereo depth estimation. The model takes an RGB stereo-image pair as input and outputs an accurate disparity map. It is designed to achieve strong zero-shot generalization across various scenarios.
Model ReleasesNvidiaFoundationStereo
FoundationStereo: A Foundation Model for Zero-Shot Stereo Depth Estimation
The model is intended for developers applying zero-shot depth to 3D perception in industrial, robotics, and smart space applications. It utilizes a mixed Transformer-CNN based architecture, incorporating features from DepthAnythingV2 and EdgeNeXt, combined with an Attentive Hybrid Cost Filtering (AHCF) cost volume and a disparity Transformer. For high-accuracy output, the initial disparity maps are refined iteratively using a convolutional GRU.
The model consists of 63 million parameters and is optimized to run on NVIDIA GPU-accelerated systems. It supports PyTorch, ONNX, and TensorRT runtimes. For deployment, it is compatible with NVIDIA Jetson devices and can be integrated with the NVIDIA Triton Inference Server. The model is available under the NVIDIA Open Model License and is ready for commercial use.
Sources
- nvidia/c-foundationstereo-s （HF: NVIDIA, 2026-09-15）