facebook/stereo-splat

StereoSplat

Model weights for StereoSplat: Metric-Scale Novel View Synthesis via Stereo-Grounded Gaussian Splatting (NeurIPS 2026).

Paper | Project Page | Code

StereoSplat predicts metric-scale 3D Gaussians from calibrated stereo pairs in a single feed-forward pass, and renders novel views and depth.

Files

File	Content	SHA-256
stereosplat_v1.safetensors	Weights; the model config is in the file metadata	e84556050f65b28367a6bab67606d742db7b4881aa202582d687ad560260184f
stereosplat_v1.json	Same config as plain JSON	8c8c57468beafb55fabbea5cd2a72efff9b994a986839f891b27033a8fdff1cb

FoundationStereo dependency

StereoSplat uses a frozen FoundationStereo backbone, which is not included here and is subject to NVIDIA's license. Get it separately:

git clone https://github.com/NVlabs/FoundationStereo third_party/FoundationStereo
git -C third_party/FoundationStereo checkout 6e8806816b533e4d13ddbb95ffa907b797060a62

Download the ViT-S model 11-33-40 from the FoundationStereo README and place it as:

checkpoints/foundation_stereo/11-33-40/
  cfg.yaml
  model_best_bp2.pth   # sha256 8d7850b9dc68d1366722a02a39745704b1db41471211be6abd93ef463727e6be

Usage

Install the code following the GitHub repository, then:

hf download facebook/stereo-splat stereosplat_v1.safetensors --local-dir checkpoints

python scripts/demo.py \
    --left path/to/left.png --right path/to/right.png \
    --intrinsics FX FY CX CY --baseline 0.25 \
    --checkpoint checkpoints/stereosplat_v1.safetensors \
    --fs-checkpoint-dir checkpoints/foundation_stereo/11-33-40 \
    --output-dir outputs/demo

The demo writes the Gaussians (.ply, standard 3DGS layout), the input-view depth, and rendered RGB and depth for novel cameras. Rendering needs a CUDA GPU.

In Python:

from huggingface_hub import hf_hub_download
from stereosplat.models.build import build_model, load_checkpoint, read_checkpoint

path = hf_hub_download("facebook/stereo-splat", "stereosplat_v1.safetensors")
config = dict(read_checkpoint(path)["config"], fs_checkpoint_path="checkpoints/foundation_stereo/11-33-40")
model = build_model(config)   # loads FoundationStereo
load_checkpoint(model, path)  # loads StereoSplat
model = model.cuda().eval()

Results

Novel view synthesis at 320 x 640 with MAC-VO poses (paper protocol: 3 consecutive stereo pairs, first and last as input, middle rendered).

Benchmark	PSNR	SSIM	LPIPS	AbsRel	delta1 (%)
Infinigen	29.41	0.954	0.116	0.033	98.56
TartanAir v2	26.25	0.944	0.101	0.058	94.95
KITTI (zero-shot)	25.20	0.941	0.101	0.044	95.98

Trained on TartanAir, Infinigen and XRStereo.

License

CC BY-NC 4.0, see LICENSE. Non-commercial use only. FoundationStereo is subject to NVIDIA's license.

Citation

@inproceedings{yugay2026stereosplat,
  title     = {StereoSplat: Metric-Scale Novel View Synthesis via Stereo-Grounded Gaussian Splatting},
  author    = {Yugay, Vladimir and Rozumny, Denis and Gevers, Theo and Vansteenkiste, Elias and Oswald, Martin R. and Luvizon, Diogo C.},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2026}
}