BLASt3R — matching model
Multi-view matching model for BLASt3R, a method for bundle-adjusted 3D reconstruction and visual SLAM from image collections and video. Code: github.com/naver/blast3r.
This is one of the two checkpoints BLASt3R needs. The other is the depth model,
naver/blast3r-depth.
Contents
config.json model class, its arguments, and the matcher slots
model.safetensors the weights
retrieval.json ASMK retrieval settings
retrieval.safetensors whitening and codebook centroids
config.json names the class and hyperparameters, so loading a checkpoint never
unpickles anything.
Usage
from huggingface_hub import snapshot_download
ckpt_match = snapshot_download('naver/blast3r-matcher')
ckpt_depth = snapshot_download('naver/blast3r-depth')
Then run reconstruction from the BLASt3R code repository:
python scripts/run_offline.py \
--images /path/to/images \
--output /path/to/output \
--ckpt_match "$CKPT_MATCH" \
--ckpt_depth "$CKPT_DEPTH" \
--mono_cam
Retrieval requires ASMK and faiss. Without
them the retrieval weights in this repository are unused and inference falls
back to a much weaker heuristic.
License
Released under the NAVER non-commercial license (LICENSE), which covers the
code and these checkpoints. Third-party subcomponents, the datasets these
weights were trained on, and the pretrained weights they were initialized from
are acknowledged in NOTICE; dataset terms pass through to the weights.
The companion depth checkpoint
(naver/blast3r-depth) was
initialized from DINOv3 weights
and carries additional terms; see the
DINOv3 License.