MoGe-2 family notice
====================

Code provenance
---------------
The MoGe encoder wrapper, convolutional neck, normal decoder, image-plane
coordinate construction, and inference graph in this directory are ported
from the official Microsoft MoGe implementation:

    https://github.com/microsoft/MoGe
    commit 925b8ed835a7a9cdb7578ba15c658a0afc969030
    License: MIT
    Copyright (c) Microsoft Corporation.

The full Microsoft MIT notice is reproduced in the repository-level
THIRD_PARTY_NOTICES.txt.

LibreYOLO omits the unused point, mask, and metric-scale heads. It reuses the
existing DINOv2 implementation already bundled for Depth Anything V2, derived
from facebookresearch/dinov2 under Apache License 2.0. No additional DINOv2
source is copied for this family.

Weight provenance
-----------------
The official ViT-S, ViT-B, and ViT-L normal checkpoints are downloaded
directly from:

    https://huggingface.co/Ruicheng/moge-2-vits-normal
    revision 679230677b4d282c6f304189a93e98e14f085902
    file model.pt
    SHA-256 79a16621928c2bf0ed04659218c55c01075e950507f40bb3332fb4c873d3e1dc
    model-card license: MIT

    https://huggingface.co/Ruicheng/moge-2-vitb-normal
    revision 54ad3a693e61907ea4633d13dec6ee682fa09419
    file model.pt
    SHA-256 16b8110e86d5dc5a849db120ca96ef3a223fd30b0c9146d1d81db504073da5f6
    model-card license: MIT

    https://huggingface.co/Ruicheng/moge-2-vitl-normal
    revision b135031bae30b5ac2ae141a0e68717795ce38340
    file model.pt
    SHA-256 280741fd09bc3f403ccff9967784c2a391b52d2c0742ae3efdb21d9f90cc1a01
    model-card license: MIT

LibreYOLO checksum-pins those objects. The converter drops only tensors
belonging to the unused point, mask, and metric-scale heads, retains all
encoder, neck, and normal-head tensors unchanged, and adds LibreYOLO checkpoint
metadata.
