Eco3R: Alignment Refocusing weights for VGGT-1B
Weights for Eco3R (paper). Eco3R is a training-free inference framework in which every frame of a long video attends to the same 31 evenly spaced frames, the Pinned Key Set. For VGGT, we also fine-tune global blocks 10 to 15, which align the frames, on samples built the same way as this inference. We call this step Alignment Refocusing.
alignment_refocusing_vggt1b.safetensors holds the 108 tensors of these six blocks (75.6M float32 parameters) and replaces them in facebook/VGGT-1B. All other weights stay unchanged.
Usage
import eco3r
from eco3r.io import load_images
model = eco3r.load_backbone("vggt") # VGGT-1B with these weights, downloaded on first use
images = load_images("path/to/frames")
rec = eco3r.reconstruct(model, images.cuda(), "vggt", score=True)
To load the weights into a VGGT model that you built yourself, call eco3r.load_refocusing(model, "alignment_refocusing_vggt1b.safetensors").
Training
We trained for 2,000 AdamW steps at a learning rate of 1e-5 on 195 ScanNet scans, the first scans of rooms that are not among the evaluation scenes. This took 3.6 hours on one A100 GPU. Each step takes a whole scan or a random contiguous half of it, pins 24 evenly spaced frames, and adds 24 random frames that attend only to the pinned ones. The loss is the camera loss of VGGT on all frames (weight 5) plus its depth loss on six of them. The training script is in the GitHub repository.
With these weights, Eco3R (VGGT) succeeds on 82% of 500-frame ScanNet videos, against 70% for the full VGGT, and runs 8.5 times as fast. Success means an absolute trajectory error below 0.2 m.
License
The weights are released under CC BY-NC 4.0 for non-commercial research, the license of VGGT-1B, from which they are fine-tuned. They were trained on ScanNet, whose terms of use also restrict use to non-commercial research.
Citation
@article{cao2026eco3r,
title = {{Eco3R}: Pinned-Key Inference for Efficient and Streamable Feed-Forward {3D} Reconstruction from Long Videos},
author = {Cao, Hongsen and Yuan, Shanxin and Jaber, Mona and Sayed, Ahmed},
journal = {arXiv preprint arXiv:2611.18099},
year = {2026}
}
Model tree for Tequila7437/Eco3R
Base model
facebook/VGGT-1B