LibreViTMatte third-party attribution and weight restriction
============================================================

PRETRAINED WEIGHTS: NON-COMMERCIAL
----------------------------------

The ``hustvl/vitmatte-small-composition-1k`` checkpoint was trained on the
Adobe Composition-1k dataset. LibreYOLO treats this pretrained checkpoint and
its converted form as subject to the Adobe Deep Image Matting Dataset License
Agreement, including its NON-COMMERCIAL restriction. Use and distribute these
weights only for non-commercial purposes unless you obtain the required rights
from Adobe. The agreement also says the Model may not be sold. Retain this
notice and the mandatory attribution below.

  Official Adobe Deep Image Matting dataset page:
  https://sites.google.com/view/deepimagematting/homepage

The agreement accompanies the dataset and is reproduced verbatim as the
`LICENSE` file in any LibreYOLO-hosted copy of this pretrained checkpoint.

The Hugging Face repository currently labels the checkpoint repository
Apache-2.0. That repository metadata does not waive or replace the terms of the
training dataset agreement. The restriction above applies to the pretrained
weights, not to the separately MIT/Apache-2.0 source port or to weights trained
independently on data for which the user has suitable rights. This notice is a
provenance and distribution safeguard, not legal advice.

Pinned source checkpoint
------------------------

  Repository: hustvl/vitmatte-small-composition-1k
  Revision:   6a58ad7646403c1df626fbd746900aec7361ea1d
  File:       model.safetensors
  Size:       103,294,572 bytes
  SHA-256:    bda9289db1bb6762d978b42d1c62ae3f34daf7497171a347a1d09657efd788cb
  URL:        https://huggingface.co/hustvl/vitmatte-small-composition-1k/blob/6a58ad7646403c1df626fbd746900aec7361ea1d/model.safetensors

Code provenance
---------------

The inference graph and preprocessing contract are ported from the audited
Apache-2.0 Hugging Face Transformers implementation at this exact commit:

  Copyright 2023 HUST-VL and The HuggingFace Inc. team.
  Copyright 2023 The HuggingFace Inc. team.
  Copyright 2023 Meta AI and The HuggingFace Inc. team.

  Repository: huggingface/transformers
  Commit:     7d6354e04794f3246bf9a0faf4fead080edeebb6
  License:    Apache License 2.0
  Source:     https://github.com/huggingface/transformers/tree/7d6354e04794f3246bf9a0faf4fead080edeebb6
  Files:
    src/transformers/models/vitmatte/modeling_vitmatte.py
    src/transformers/models/vitmatte/image_processing_vitmatte.py
    src/transformers/models/vitdet/modeling_vitdet.py

ViTMatte's architecture and checkpoint lineage originate with HUST-VL:

  Repository: hustvl/ViTMatte
  Commit:     f48208ceb0d172ab0c1c528bba33cc1159596407
  License:    MIT
  Copyright (c) 2023 Hust Vision Lab
  Source:     https://github.com/hustvl/ViTMatte/tree/f48208ceb0d172ab0c1c528bba33cc1159596407

No HUST training-stack code was copied or adapted. LibreViTMatte does not use
HUST's easydict, detectron2, timm, or fairscale dependencies. The family-local
inference implementation loads the pinned Transformers-format state dict
strictly and is externally checked against the Transformers reference.

Required attribution
--------------------

Adobe's agreement requires any use of or reference to the Dataset, and any
permitted distribution of images or the Model, to include this exact
attribution:

  Ning Xu, Brian Price, Scott Cohen, and Thomas Huang. Deep Image Matting.
  Proceedings of the Conference on Computer Vision and Pattern Recognition,
  2017.

Optional model-paper citation
-----------------------------

  Jingfeng Yao, Xinggang Wang, Shusheng Yang, and Baoyuan Wang,
  "ViTMatte: Boosting Image Matting with Pretrained Plain Vision Transformers,"
  Information Fusion, 2024.

LibreYOLO integration
---------------------

The family wrapper, strict trimap validation, postprocessing, checkpoint
converter, parity harness, and tests are original LibreYOLO work and are
MIT-licensed with the repository. Apache-2.0 notices for the ported graph are
preserved by this file and the repository's license surfaces.
