MASt3R ViT-Large (SafeTensors)

Official MASt3R (Matching And Stereo 3D Reconstruction) checkpoint converted to SafeTensors format for efficient inference.

Files

File	Description	Size
`unified.safetensors`	Main model (encoder + decoder)	~1.3 GB
`retrieval.safetensors`	Retrieval head for SfM	~2 MB
`codebook.pkl`	Codebook for image retrieval	~256 MB

Model Details

Architecture: ViT-Large/14 encoder + ViT-Base decoder
Resolution: 512x512 (native), supports 512x384, 512x336, 512x288
Head: CatMLP + DPT
Precision: FP16
Parameters: ~1B total

Capabilities

Dense 3D Reconstruction: Per-pixel 3D point prediction
Feature Matching: Dense descriptors for stereo matching
Image Retrieval: Codebook-based retrieval for SfM pipelines
Multi-view Reconstruction: Full 3D scene reconstruction from image collections

Usage

With mast3r-runtime

from mast3r_runtime import load_model

model = load_model("aedelon/mast3r-vit-large")
points, descriptors = model.forward(image1, image2)

Direct loading

from huggingface_hub import hf_hub_download
from safetensors.torch import load_file
import pickle

# Load main model
unified = load_file(hf_hub_download("Aedelon/mast3r-vit-large-fp16", "unified.safetensors"))

# Load retrieval head
retrieval = load_file(hf_hub_download("Aedelon/mast3r-vit-large-fp16", "retrieval.safetensors"))

# Load codebook (pickle)
codebook_path = hf_hub_download("Aedelon/mast3r-vit-large-fp16", "codebook.pkl")
with open(codebook_path, "rb") as f:
    codebook = pickle.load(f)

With original MASt3R

from mast3r.model import AsymmetricMASt3R
from safetensors.torch import load_file

weights = load_file("unified.safetensors")
model = AsymmetricMASt3R()
model.load_state_dict(weights)

Output Format

# Per image pair:
{
    "pts3d_1": Tensor[B, H, W, 3],      # 3D points from view 1
    "pts3d_2": Tensor[B, H, W, 3],      # 3D points from view 2
    "desc_1": Tensor[B, H, W, D],       # Descriptors from view 1
    "desc_2": Tensor[B, H, W, D],       # Descriptors from view 2
    "conf_1": Tensor[B, H, W],          # Confidence map view 1
    "conf_2": Tensor[B, H, W],          # Confidence map view 2
}

Citation

@article{leroy2024mast3r,
    title={Grounding Image Matching in 3D with MASt3R},
    author={Leroy, Vincent and Cabon, Yohann and Revaud, J{\'e}r{\^o}me},
    journal={arXiv preprint arXiv:2406.09756},
    year={2024}
}

@inproceedings{wang2024dust3r,
    title={DUSt3R: Geometric 3D Vision Made Easy},
    author={Wang, Shuzhe and Leroy, Vincent and Cabon, Yohann and Chidlovskii, Boris and Revaud, J{\'e}r{\^o}me},
    booktitle={CVPR},
    year={2024}
}