R3 — Training code
==================

This directory (`R3/training/`) contains the training code used to produce the
released R3 checkpoints. It is provided for NON-COMMERCIAL RESEARCH USE ONLY,
because it incorporates and adapts third-party components released under
non-commercial licenses (see below).

Third-party components and attribution
--------------------------------------

- DUSt3R / CroCo (NAVER Corporation)
  Portions of `dust3r/` and `croco/` are adapted from the DUSt3R and CroCo
  projects. Original code: Copyright (C) NAVER Corporation, licensed under
  CC BY-NC-SA 4.0 (non-commercial use only). Original license headers are
  retained in the corresponding source files.
  https://github.com/naver/dust3r , https://github.com/naver/croco

- CUT3R
  The multi-view dataset loaders under `datasets/` are adapted
  from the CUT3R project. Datasets must be preprocessed following CUT3R's
  instructions; no dataset data is distributed here.
  https://github.com/CUT3R/CUT3R

- VGGT
  Camera, pose, and geometry utilities are adapted from or inspired by VGGT
  (Visual Geometry Grounded Transformer) components.
  https://github.com/facebookresearch/vggt

- Depth Anything 3
  Training distills from and initializes the backbone with publicly released
  Depth Anything 3 weights (https://huggingface.co/depth-anything). Those
  weights are governed by their own upstream license.

Summary
-------

Because of the DUSt3R/CroCo/CUT3R/VGGT lineage, the R3 training code and any
models trained with it are intended for non-commercial research purposes only.
This NOTICE does not modify the licenses of the upstream projects; consult each
upstream repository for authoritative terms.