Aedelon/dunemast3r-models-fp16

MASt3R & DUNE Checkpoints (SafeTensors)

Pre-trained checkpoints for MASt3R (Matching And Stereo 3D Reconstruction) with DUNE (Dense UNconstrained Estimation) models and DUNEMASt3R models, converted to SafeTensors format for efficient C++/embedded inference.

Models

Model	Resolution	Encoder	Size	Use Case
dune_vit_small_336	336x336	ViT-S/14	~1.3 GB	Real-time drone/embedded
dune_vit_small_448	448x448	ViT-S/14	~1.3 GB	Fast inference
dune_vit_base_336	336x336	ViT-B/14	~1.7 GB	Balanced speed/quality
dune_vit_base_448	448x448	ViT-B/14	~1.7 GB	High quality

Architecture

• Encoder: DINOv2-based Vision Transformer (DUNE-trained)
• Decoder: MASt3R decoder with CatMLP+DPT heads
• Outputs: Dense 3D points + descriptors for matching

Image Pair → DUNE Encoder → MASt3R Decoder → 3D Points + Descriptors

Usage

With mast3r-runtime (recommended)

pip install mast3r-runtime

# Download and convert
mast3r-runtime download dune_vit_small_336
mast3r-runtime convert dune_vit_small_336 --dtype fp16

Direct download

from huggingface_hub import hf_hub_download

# Download encoder
encoder = hf_hub_download(
    repo_id="Aedelon/dunemast3r-models-fp16",
    filename="dune_vit_small_336/encoder.safetensors"
)

# Download decoder
decoder = hf_hub_download(
    repo_id="Aedelon/dunemast3r-models-fp16",
    filename="dune_vit_small_336/decoder.safetensors"
)

Credits & Acknowledgments

These models are converted from the original checkpoints released by Naver Labs Europe.

MASt3R

Grounding Image Matching in 3D with MASt3R Vincent Leroy, Yohann Cabon, Jérôme Revaud arXiv 2024

@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}
}

DUNE

DUNE: Dense UNconstrained Estimation for 3D Vision Vincent Leroy, Yohann Cabon, Jérôme Revaud CVPR 2025

@inproceedings{leroy2025dune,
  title={DUNE: Dense UNconstrained Estimation for 3D Vision},
  author={Leroy, Vincent and Cabon, Yohann and Revaud, J{\'e}r{\^o}me},
  booktitle={CVPR},
  year={2025}
}

DUSt3R

DUSt3R: Geometric 3D Vision Made Easy Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, Jérôme Revaud CVPR 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}
}

Original Repositories

• naver/mast3r - MASt3R official implementation
• naver/dune - DUNE official implementation
• naver/dust3r - DUSt3R official implementation

License

The model weights are released under CC BY-NC-SA 4.0 (Creative Commons Attribution-NonCommercial-ShareAlike 4.0).

• Attribution: Credit Naver Labs Europe
• NonCommercial: No commercial use without permission
• ShareAlike: Derivatives must use same license

For commercial licensing, contact Naver Labs Europe.

Converted by

Delanoe Pirard / Aedelon - mast3r-runtime

SafeTensors conversion for embedded/C++ inference (Apache 2.0 for runtime code).