This repository includes code derived from, or inspired by, the
following open-source projects. Each upstream is listed with its
license and the LibreYOLO module(s) that port from it.

--------------------------------------------------------------------
COCO API / pycocotools
--------------------------------------------------------------------
Source: https://github.com/ppwwyyxx/cocoapi
Commit: ac87f5077ad6b8864c2dc5e93d14cae62d1db05a
Version: pycocotools 2.0.11
License: FreeBSD (BSD-2-Clause)
Copyright (c) 2014, Piotr Dollar and Tsung-Yi Lin
Used for: the configurable maxDets metric extraction in
          libreyolo/validation/coco_evaluator.py, adapted from COCOeval's
          accumulated precision/recall tensor layout and summary logic.

Copyright (c) 2014, Piotr Dollar and Tsung-Yi Lin

All rights reserved.

Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are met:
1. Redistributions of source code must retain the above copyright notice, this
   list of conditions and the following disclaimer.
2. Redistributions in binary form must reproduce the above copyright notice,
   this list of conditions and the following disclaimer in the documentation
   and/or other materials provided with the distribution.

THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

The views and conclusions contained in the software and documentation are those
of the authors and should not be interpreted as representing official policies,
either expressed or implied, of the FreeBSD Project.

--------------------------------------------------------------------
SAHI
--------------------------------------------------------------------
Source: https://github.com/obss/sahi
License: MIT
Copyright (c) 2020 obss
Used for: slicing-aided hyper inference utilities.

MIT License

Copyright (c) 2020 obss

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.

--------------------------------------------------------------------
CLIP / OpenCLIP (OpenAI; LAION / ML Foundations)
--------------------------------------------------------------------
Source: https://github.com/openai/CLIP, https://github.com/mlfoundations/open_clip
License: MIT
Copyright (c) 2021 OpenAI; (c) 2012-2021 OpenCLIP authors
Used for: the LibreCLIP family (libreyolo/models/clip/). The byte-pair-encoding
text tokenizer (libreyolo/models/clip/tokenizer.py) and the bundled BPE merge
table (libreyolo/models/clip/bpe_simple_vocab_16e6.txt.gz) are vendored from the
CLIP / open_clip tokenizer. The image/text towers are a clean-room native torch
re-implementation of the standard CLIP architecture (no open_clip at runtime).

The shipped LibreCLIP weights are converted from OpenCLIP LAION-2B checkpoints,
which are MIT-redistributable. NOTE: the LAION-2B training data has a documented
CSAM-content history (Stanford, 2023); use Re-LAION-derived weights. See
libreyolo/models/clip/NOTICE.md for the full data-provenance note.

MIT License

Copyright (c) 2021 OpenAI

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.

--------------------------------------------------------------------
SigLIP / SigLIP 2 (Google; Hugging Face Transformers)
--------------------------------------------------------------------
Source: https://github.com/google-research/big_vision,
        https://github.com/huggingface/transformers (models/siglip, models/siglip2)
License: Apache License 2.0
Copyright (c) Google LLC; (c) The HuggingFace Inc. team.
Used for: the LibreSigLIP2 family (libreyolo/models/siglip2/). The image/text
towers (libreyolo/models/siglip2/nn.py) are a clean-room native torch
re-implementation of the SigLIP architecture, structured to match the
transformers reference implementation (no transformers at runtime). The
multilingual SentencePiece tokenizer model
(libreyolo/models/siglip2/siglip2_tokenizer.model, Gemma vocabulary) is shipped
verbatim from the Apache-2.0 google/siglip2-* Hugging Face release.

The shipped LibreSigLIP2 weights are converted from the Apache-2.0
google/siglip2-base-patch16-256 and google/siglip2-so400m-patch14-384
checkpoints (state-dict metadata wrap only; learned parameters unchanged). See
libreyolo/models/siglip2/NOTICE.md.

--------------------------------------------------------------------
YOLOX (Megvii-BaseDetection)
--------------------------------------------------------------------
Source: https://github.com/Megvii-BaseDetection/YOLOX
License: Apache License 2.0
Copyright (c) 2021-2022 Megvii Inc. All rights reserved.
Used for: YOLOX model family (libreyolo/models/yolox/), EMA helper
          (libreyolo/training/ema.py), augmentation pipeline
          (libreyolo/training/augment.py), and the SimOTA training loss
          adapted for the YOLOv7 family (libreyolo/models/yolo7/loss.py:
          imports bboxes_iou/IoULoss from the yolox modules and adapts the
          get_assignments/get_geometry_constraint/simota_matching/get_losses
          assignment logic to the v7 anchor head).

--------------------------------------------------------------------
YOLO (MultimediaTechLab/YOLO)
--------------------------------------------------------------------
Source: https://github.com/MultimediaTechLab/YOLO
License: MIT
Copyright (c) 2024 Kin-Yiu Wong and Hao-Tang Tsui
Used for: YOLO9 model family (libreyolo/models/yolo9/ and
          libreyolo/models/yolo9_e2e/): the architecture blocks and
          detection head in nn.py and the loss port in loss.py.
          Also the YOLOv7 family (libreyolo/models/yolo7/): the architecture
          (net.py, blocks.py) and the Anc2Box anchor decode reproduced in
          postprocess/yolo7.py and mirrored by the training loss (loss.py:
          _decode). Upstream ships no v7 training loss, so training assignment
          is adapted from YOLOX (see the YOLOX entry above), not from here.

--------------------------------------------------------------------
RepVGG (DingXiaoH)
--------------------------------------------------------------------
Source: https://github.com/DingXiaoH/RepVGG
License: MIT
Copyright (c) 2020 DingXiaoH
Used for: RepConv fuse / re-parameterization logic
          (libreyolo/models/yolo9/nn.py RepConvN.fuse_convs).

--------------------------------------------------------------------
mmdetection (OpenMMLab)
--------------------------------------------------------------------
Source: https://github.com/open-mmlab/mmdetection
Commit: cfd5d3a985b0249de009b67d04f37263e11cdf3d
License: Apache License 2.0
Copyright (c) OpenMMLab. All rights reserved.
Used for: RTMDet model family (libreyolo/models/rtmdet/): architecture
          port in nn.py, including the RTMDet-Ins head and mask decoder;
          QualityFocalLoss, GIoULoss,
          DynamicSoftLabelAssigner and MlvlPointGenerator in loss.py.
          Published RTMDet and RTMDet-Ins COCO weights were trained with
          mmdetection.

--------------------------------------------------------------------
mmsegmentation (OpenMMLab)
--------------------------------------------------------------------
Source: https://github.com/open-mmlab/mmsegmentation
        (mmseg/datasets/transforms/transforms.py)
License: Apache License 2.0
Copyright (c) OpenMMLab. All rights reserved.
Used for: the dense random-crop sampling used by semantic training
          (libreyolo/data/semantic_dataset.py: the optional
          ``resize_crop`` mode and its ``crop_cat_max_ratio`` retry
          loop). The recipe -- resize the short side, pad, then re-sample
          a crop up to 10 times until no single class exceeds
          ``cat_max_ratio`` of its non-ignored pixels -- is derived from
          mmsegmentation's ``RandomCrop`` (``cat_max_ratio``), as are the
          SegFormer ADE20K training hyper-parameters (decode-head LR
          multiplier, no weight decay on norms and the Mix-FFN positional
          conv, scale jitter 0.5-2.0). NOT derived from NVIDIA's
          NVlabs/SegFormer fork of mmseg 0.x, which is non-commercial.

--------------------------------------------------------------------
PicoDet (PaddleDetection / Picodet_Pytorch)
--------------------------------------------------------------------
Source: https://github.com/Bo396543018/Picodet_Pytorch (direct source,
        a PyTorch port built on mmdetection), from
        https://github.com/PaddlePaddle/PaddleDetection (original)
License: Apache License 2.0 (all of Picodet_Pytorch, PaddleDetection,
         and mmdetection)
Copyright (c) PaddlePaddle Authors; OpenMMLab.
Used for: PICODET model family (libreyolo/models/picodet/).

--------------------------------------------------------------------
PIDNet (XuJiacong)
--------------------------------------------------------------------
Source: https://github.com/XuJiacong/PIDNet
License: MIT
Copyright (c) 2022 Jiacong Xu
Used for: PIDNet semantic segmentation family
          (libreyolo/models/pidnet/). Converted Cityscapes weights are
          MIT-licensed PIDNet weights; the Cityscapes dataset itself is not
          redistributed by LibreYOLO.

--------------------------------------------------------------------
SuperGradients / YOLO-NAS
--------------------------------------------------------------------
Source: https://github.com/Deci-AI/super-gradients
License: Apache License 2.0
Copyright (c) 2021-2024 Deci AI
Used for: YOLO-NAS model family and pose training references
          (libreyolo/models/yolonas/). YOLO-NAS source code is
          Apache-2.0; published pretrained YOLO-NAS weights may have
          separate non-commercial terms and are not bundled here.

--------------------------------------------------------------------
EdgeCrafter
--------------------------------------------------------------------
Source: https://github.com/EC-codehub/EdgeCrafter
License: Apache License 2.0
Used for: EC model family detection, segmentation, and pose architecture
          references (libreyolo/models/ec/).

--------------------------------------------------------------------
D-FINE-seg (ArgoHA)
--------------------------------------------------------------------
Source: https://github.com/ArgoHA/D-FINE-seg
License: Apache License 2.0
Copyright (c) 2026 The D-FINE-seg Authors. All Rights Reserved.
Used for: D-FINE instance-segmentation mask decoder/head, mask matching,
          mask loss, and postprocess references
          (libreyolo/models/dfine/, libreyolo/postprocess/dfine.py).
          The repository maintainer approved reuse with attribution in
          ArgoHA/D-FINE-seg#70.

--------------------------------------------------------------------
Dome-DETR (RicePasteM)
-------------------------------------------------------------------
Source: https://github.com/RicePasteM/Dome-DETR
Commit: 2dde3bc1946a3e9fad9abd0612b59fc39bd6b861
License: Apache License 2.0
Copyright (c) 2025 The Dome-DETR Authors. All Rights Reserved.
Used for: the Dome-DETR tiny-object detection family
          (libreyolo/models/domedetr/, libreyolo/postprocess/domedetr.py,
          weights/convert_domedetr_weights.py, weights/parity_domedetr.py).
          Specifically the DeFE density head, the MWAS window processor,
          and PAQI query initialisation with its density-adaptive NMS; the
          static-shape MWAS formulation follows the same repository's
          onnx-export branch. Upstream is itself modified from D-FINE
          (Apache-2.0), and the parts Dome-DETR shares with it are imported
          from libreyolo/models/dfine/ rather than re-vendored.
          The upstream LICENSE leaves the copyright line as the
          "[yyyy] [name of copyright owner]" placeholder; the attribution
          above comes from the upstream source headers.

-------------------------------------------------------------------
RT-DETR (lyuwenyu)
--------------------------------------------------------------------
Source: https://github.com/lyuwenyu/RT-DETR
License: Apache License 2.0
Copyright (c) 2023 lyuwenyu
Used for: RT-DETR model family (libreyolo/models/rtdetr/) including
          backbone, neck, decoder, loss, and denoising modules. The
          HGNetv2 backbone (libreyolo/models/rtdetr/hgnetv2.py) is
          ported from rtdetrv2_pytorch/src/nn/backbone/hgnetv2.py.

--------------------------------------------------------------------
RT-DETRv2 OBB (RicePasteM / RiO-DETR)
--------------------------------------------------------------------
Source: https://github.com/RicePasteM/RiO-DETR
Pinned commit: 22d5232a4e0df6ac4bc26ed1c8aac8b4060449c7
License: Apache License 2.0
Used for: the inference-only RT-DETRv2 OBB encoder and decoder in
          libreyolo/models/rtdetrv2/obb_encoder.py and obb_decoder.py.
          The port adapts engine/rtv4/hybrid_encoder.py and
          engine/rtv4/rtdetrv2_obb_decoder.py. The official DOTA 1.0
          checkpoints were validated from RicePasteM/RT-DETR-OBB at
          revision f376e9dcedfb9a47a21ac71ef61ad99f8b545698. Converted
          mirrors are published as LibreYOLO/LibreRTDETRv2{n,s,m,l,x}-obb;
          every repository carries the upstream Apache License 2.0 and
          attribution notice. Users may also convert a locally supplied
          official checkpoint with weights/convert_rtdetrv2_weights.py.

--------------------------------------------------------------------
RF-DETR (Roboflow)
--------------------------------------------------------------------
Source: https://github.com/roboflow/rf-detr
License: Apache License 2.0
Copyright (c) 2024-2025 Roboflow, Inc.
Used for: RF-DETR model family (libreyolo/models/rfdetr/), LoRA
          adapter recipe helpers (libreyolo/training/lora.py), and
          COCO evaluation glue (libreyolo/data/yolo_coco_api.py).
          Also the GroupPose-style keypoint/pose head, dual-projector,
          keypoint decoder token stream, probabilistic (Cholesky)
          keypoint regression, and keypoint postprocess ported from
          RF-DETR v1.8.0 into libreyolo/models/rfdetr/. The published
          RF-DETR keypoint preview weights (Apache-2.0, COCO person
          pretrained) are redistributed with attribution.

--------------------------------------------------------------------
DINOv2 (Meta AI / facebookresearch)
--------------------------------------------------------------------
Source: https://github.com/facebookresearch/dinov2
License: Apache License 2.0
Copyright (c) Meta Platforms, Inc. and affiliates.
Used for: vision transformer backbone consumed by RF-DETR. The local
          DINOv2 implementation lives at libreyolo/models/rfdetr/dinov2.py.

--------------------------------------------------------------------
HuggingFace Transformers
--------------------------------------------------------------------
Source: https://github.com/huggingface/transformers
License: Apache License 2.0
Copyright 2022-2024 The HuggingFace Team. All Rights Reserved.
Used for: DINOv2-with-Registers reference implementation that
          libreyolo/models/rfdetr/dinov2.py adapts to add windowed
          self-attention. Also a runtime dependency loaded via
          AutoBackbone for the non-windowed DinoV2 path. The Deformable
          DETR checkpoint key mapping in
          libreyolo/models/deformable_detr/conversion.py is adapted from
          Transformers commit
          4a224b1e2182d1f8f27d1d76fb8de6ab40b7ff62.

--------------------------------------------------------------------
EoMT (Mobile Perception Systems Lab at TU/e)
--------------------------------------------------------------------
Source: https://github.com/tue-mps/eomt
License: MIT
Copyright (c) 2025 Mobile Perception Systems Lab at TU/e
Citation: Kerssies, T., Cavagnero, N., Hermans, A., Norouzi, N.,
          Averta, G., Leibe, B., Dubbelman, G., and de Geus, D.
          "Your ViT is Secretly an Image Segmentation Model." CVPR 2025.
Used for: LibreEoMT semantic, instance, and panoptic segmentation family
          (libreyolo/models/eomt/). Runtime execution uses the Apache-2.0
          Hugging Face Transformers EoMT implementation with converted
          MIT-licensed DINOv2 EoMT weights:
            - ADE20K 150-class semantic (l, 512px)
            - COCO 80-class instance segmentation (l, 640px and 1280px)
            - COCO 133-class panoptic, task="panoptic" (s/b/l, 640px)

NOTE - code vs. weights: LibreYOLO ships only DINOv2-based EoMT checkpoints.
DINOv3 EoMT variants are excluded because they depend on gated
non-commercial DINOv3 weights.

--------------------------------------------------------------------
SegFormer / HuggingFace Transformers
--------------------------------------------------------------------
Source: https://github.com/huggingface/transformers
        (models/segformer/{configuration_segformer.py, modeling_segformer.py})
License: Apache License 2.0
Copyright 2021 NVIDIA and The HuggingFace Inc. team. All rights reserved.
Citation: Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J. M., and
          Luo, P. "SegFormer: Simple and Efficient Design for Semantic
          Segmentation with Transformers." NeurIPS 2021.
Used for: LibreSegformer semantic segmentation family
          (libreyolo/models/segformer/). The MiT encoder (overlap patch
          embeddings, efficient self-attention with spatial reduction,
          Mix-FFN) and the all-MLP decode head are a native
          reimplementation behaviorally derived from HuggingFace
          Transformers' Apache-2.0 modeling_segformer.py, NOT from
          NVIDIA's original NVlabs/SegFormer repository (NVIDIA Source
          Code License, non-commercial/research-only — never read or
          derived from). LibreSegformer has no runtime dependency on
          the transformers package.

NOTE - code vs. weights: the CODE above is Apache-2.0, but the pretrained
WEIGHTS are NOT. LibreSegformer{b0..b5}-sem are converted from NVIDIA's
ADE20K SegFormer checkpoints (nvidia/segformer-b0..b5-finetuned-ade-*),
released under the NVIDIA Source Code License:

    https://github.com/NVlabs/SegFormer/blob/master/LICENSE

That license permits redistribution provided a complete copy of the
license accompanies the weights and attribution notices are retained,
but it limits USE to non-commercial "research or evaluation purposes
only", and Section 3.2 carries the limit into every derivative work.
These weights are therefore NON-COMMERCIAL ONLY and are not covered by
LibreYOLO's permissive license; the restriction binds end users, not
just LibreYOLO. A notice is printed before every auto-download.
Conversion is a key remapping only (weights/convert_segformer_weights.py);
learned parameters are NVIDIA's, unchanged. Models the user trains from
scratch carry no such restriction.

--------------------------------------------------------------------
Grounding DINO (IDEA-Research)
--------------------------------------------------------------------
Source: https://github.com/IDEA-Research/GroundingDINO
License: Apache License 2.0
Copyright (c) 2023 IDEA-Research
Citation: Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J.,
          Li, C., Yang, J., Su, H., Zhu, J., and Zhang, L. "Grounding
          DINO: Marrying DINO with Grounded Pre-Training for Open-Set
          Object Detection." ECCV 2024.
Used for: LibreGroundingDINO open-vocabulary detector. The shipped path
          (libreyolo/models/openvocab/grounding_dino.py) runs through the
          Apache-2.0 Hugging Face Transformers GroundingDinoForObjectDetection
          implementation. A native clean-room port derived from the same
          Apache-2.0 transformers reference also lives at
          libreyolo/models/grounding_dino/. Weights are rehosted at
          LibreYOLO/LibreGroundingDINOt and LibreYOLO/LibreGroundingDINOb.

--------------------------------------------------------------------
OWLv2 / OWL-ViT (Google Research)
--------------------------------------------------------------------
Source: https://github.com/google-research/scenic (OWL-ViT / OWLv2)
License: Apache License 2.0
Copyright (c) 2023 Google LLC
Citation: Minderer, M., Gritsenko, A., and Houlsby, N. "Scaling
          Open-Vocabulary Object Detection." NeurIPS 2023.
Used for: LibreOWLv2 open-vocabulary detector. The shipped path
          (libreyolo/models/openvocab/owlv2.py) runs through the Apache-2.0
          Hugging Face Transformers Owlv2ForObjectDetection implementation.
          A native clean-room port derived from the same Apache-2.0
          transformers reference also lives at libreyolo/models/owlv2/.
          Weights are rehosted at LibreYOLO/LibreOWLv2b16 and
          LibreYOLO/LibreOWLv2l14.

--------------------------------------------------------------------
OMDet-Turbo (Om Research Lab / Hugging Face Transformers)
--------------------------------------------------------------------
Architecture: https://github.com/om-ai-lab/OmDet
Reference implementation: https://github.com/huggingface/transformers
Path: src/transformers/models/omdet_turbo/
License: Apache License 2.0
Copyright 2024 Om Research Lab and The HuggingFace Inc. team.
Used for: LibreOMDetTurbo open-vocabulary detection. The adapter at
          libreyolo/models/openvocab/omdet_turbo.py calls the Transformers
          OmDetTurboForObjectDetection implementation, its processor, and its
          post-processing. No OMDet-Turbo model source is vendored.

Weights: omlab/omdet-turbo-swin-tiny-hf revision
         7fe93cecfb770c4d76cf71163956221249cab566, Apache-2.0, mirrored
         without learned-parameter changes at LibreYOLO/LibreOMDetTurbot.

--------------------------------------------------------------------
OV-DEIM (wleilei)
--------------------------------------------------------------------
Source: https://github.com/wleilei/OV-DEIM
License: Apache License 2.0 (code); CC BY-NC 4.0 (released checkpoints,
         per upstream MODEL_LICENSE)
Citation: arXiv 2603.07022, "OV-DEIM: Real-time DETR-Style
          Open-Vocabulary Object Detection with GridSynthetic
          Augmentation."
Used for: LibreOVDEIM open-vocabulary detector, a native port vendored
          at libreyolo/models/openvocab/ovdeim/ under Apache-2.0
          (RT-DETR / DEIMv2 lineage). Converted S/M/L detector weights
          are rehosted at LibreYOLO/LibreOVDEIM{s,m,l} under
          CC BY-NC 4.0 with attribution, as the upstream MODEL_LICENSE
          permits. The text tower is the MobileCLIP-B(LT) text
          transformer (apple/MobileCLIP-B-LT-OpenCLIP); its license
          text and attribution notice ship in the weight repositories.
          Licensing was confirmed by the upstream author
          (wleilei/OV-DEIM#4); see docs/provenance/ov_deim.md.

--------------------------------------------------------------------
LW-DETR (Atten4Vis / Baidu)
--------------------------------------------------------------------
Source: https://github.com/Atten4Vis/LW-DETR
License: Apache License 2.0
Copyright (c) 2024 Baidu. All Rights Reserved.
Used for: (1) the standalone LibreLWDETR family — plain-ViT
          encoder, multi-scale projector, deformable DETR decoder,
          and postprocess (libreyolo/models/lwdetr/,
          libreyolo/postprocess/lwdetr.py); and (2) backbone,
          transformer, matcher, loss, postprocess, and tensor
          utilities consumed by RF-DETR, which forked and modified
          this architecture
          (libreyolo/models/rfdetr/{backbone,transformer,matcher,
          loss,lwdetr,tensors,box_ops}.py).
          Weights for the standalone family are converted from the
          upstream Apache-2.0 COCO releases (huggingface.co/xbsu/
          LW-DETR) and rehosted at LibreYOLO/LibreLWDETR{t,s,m,l,x};
          see docs/provenance/lwdetr.md.

--------------------------------------------------------------------
IDEA-Research DINO
--------------------------------------------------------------------
Source: https://github.com/IDEA-Research/DINO
Commit: d84a491d41898b3befd8294d1cf2614661fc0953
License: Apache License 2.0
Copyright (c) 2022 IDEA. All Rights Reserved.
Used for: the inference-only LibreDINO-DETR family
          (libreyolo/models/dinodetr/ and
          libreyolo/postprocess/dinodetr.py). The three official checkpoints
          are converted from the authors' Google Drive releases and rehosted
          at LibreYOLO/LibreDINODETR{r50,r50s5,swinl}; see
          docs/provenance/dinodetr.md. The releasing repository declares
          Apache-2.0, but the checkpoint files carry no standalone license
          file or metadata; every mirror ships that repository license and a
          notice explaining the redistribution basis.

--------------------------------------------------------------------
DAB-DETR and DN-DETR (IDEA-Research)
--------------------------------------------------------------------
Sources: https://github.com/IDEA-Research/DAB-DETR
         https://github.com/IDEA-Research/DN-DETR
License: Apache License 2.0
Used for: architectural predecessors identified by the pinned DINO source;
          the LibreDINO-DETR port is made from DINO itself rather than from
          separate copies of these repositories.

--------------------------------------------------------------------
Conditional DETR (Atten4Vis / Microsoft)
--------------------------------------------------------------------
Source: https://github.com/Atten4Vis/ConditionalDETR
License: Apache License 2.0
Copyright (c) 2021 Microsoft. All Rights Reserved.
Used for: position-encoding, transformer, matcher, and loss
          building blocks reused by RF-DETR via LW-DETR
          (libreyolo/models/rfdetr/{backbone,transformer,matcher,
          loss,lwdetr,box_ops}.py), plus inherited transformer and backbone
          portions of LibreDINO-DETR (libreyolo/models/dinodetr/).

--------------------------------------------------------------------
DETR (facebookresearch / Meta)
--------------------------------------------------------------------
Source: https://github.com/facebookresearch/detr
Commit: 29901c51d7fe8712168b8d0d64351170bc0f83e0
License: Apache License 2.0
Copyright (c) Facebook, Inc. and its affiliates.
Used for: the standalone original DETR architecture and inference utilities
          (libreyolo/models/detr/, libreyolo/postprocess/detr.py), plus
          NestedTensor, position-encoding, matcher, set-criterion, and box
          utilities reused by RF-DETR via LW-DETR
          (libreyolo/models/rfdetr/{backbone,transformer,matcher,
          loss,lwdetr,tensors,box_ops}.py), and the NestedTensor,
          position-encoding, and ResNet-backbone portions of the standalone
          Deformable DETR port (libreyolo/models/deformable_detr/common.py),
          with inherited backbone and transformer portions in
          libreyolo/models/dinodetr/.

--------------------------------------------------------------------
PyTorch torch.nn.Transformer
--------------------------------------------------------------------
Source: https://github.com/pytorch/pytorch
License: BSD-3-Clause
Copyright (c) 2016-2026 Facebook, Inc. and its affiliates.
Used for: facebookresearch/detr models/transformer.py states that it is a
          modified copy of torch.nn.Transformer. LibreDETR ports that
          Apache-2.0 DETR module and retains the inherited PyTorch lineage
          (libreyolo/models/detr/nn.py).

--------------------------------------------------------------------
Deformable DETR (fundamentalvision / SenseTime)
--------------------------------------------------------------------
Source: https://github.com/fundamentalvision/Deformable-DETR
        commit 11169a60c33333af00a4849f1808023eba96a931
License: Apache License 2.0
Copyright (c) 2020 SenseTime. All Rights Reserved.
Used for: the standalone inference-only LibreDeformableDETR family
          (libreyolo/models/deformable_detr/ and
          libreyolo/postprocess/deformable_detr.py), plus multi-scale
          deformable attention reused by RF-DETR
          (libreyolo/models/rfdetr/transformer.py: MSDeformAttn,
          ms_deform_attn_core_pytorch) and LibreDINO-DETR
          (libreyolo/models/dinodetr/). The five standalone checkpoints are
          converted from the Apache-2.0 SenseTime Hugging Face mirrors and
          rehosted at LibreYOLO/LibreDeformableDETR{r50ss,r50ssdc5,r50,
          r50refine,r50twostage}; see docs/provenance/deformable_detr.md.
          The autograd bridge in
          libreyolo/kernels/attention/ms_deform_attn.py follows the upstream
          MSDeformAttnFunction interface; the compiled CUDA kernel itself is
          NOT vendored. It is fetched at runtime from the Apache-2.0 Hub
          repository kernels-community/deformable-detr, only when the
          optional `kernels` package (the libreyolo[hub-kernels] extra) is
          installed; LIBREYOLO_HUB_KERNELS=0 disables it.

--------------------------------------------------------------------
Swin Transformer (Microsoft)
--------------------------------------------------------------------
Source: https://github.com/microsoft/Swin-Transformer
License: MIT
Copyright (c) 2021 Microsoft.
Used for: the Swin-L backbone ported through DINO for
          LibreDINODETRswinl (libreyolo/models/dinodetr/swin.py).

--------------------------------------------------------------------
ViTDet (facebookresearch detectron2)
--------------------------------------------------------------------
Source: https://github.com/facebookresearch/detectron2/tree/main/projects/ViTDet
License: Apache License 2.0
Copyright (c) Facebook, Inc. and its affiliates.
Used for: MultiScaleProjector / SimpleProjector primitives reused by
          RF-DETR (libreyolo/models/rfdetr/backbone.py).

--------------------------------------------------------------------
PaddleClas (PaddlePaddle)
--------------------------------------------------------------------
Source: https://github.com/PaddlePaddle/PaddleClas
License: Apache License 2.0
Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
Used for: ResNet_vd pretrained classification backbones loaded by
          RT-DETR (libreyolo/models/rtdetr/backbone.py downloads
          ResNet{18,34,50,101}_vd weights that originate here).

--------------------------------------------------------------------
MiDaS (Intel ISL)
--------------------------------------------------------------------
Source: https://github.com/isl-org/MiDaS
Pinned commit: 454597711a62eabcbf7d1e89f3fb9f569051ac9b
License: MIT
Copyright (c) 2019 Intel ISL (Intel Intelligent Systems Lab)
Used for: the native inference-only MiDaS v2.1 Small and DPT-Large
          relative-depth family in libreyolo/models/midas/. Module names and
          arithmetic follow the pinned source so the official release state
          dicts load unchanged and produce bit-exact outputs. LibreYOLO adds
          its checkpoint metadata, depth-task result, validation, and fixed
          export contracts.

NOTE - weights: LibreYOLO does not bundle or rehost the official release
checkpoints. They are downloaded directly from isl-org/MiDaS and SHA-256
verified before safe loading. Although the releasing repository is MIT, the
commercial-use clearance of the mixed training datasets has not been
established as required for hosted depth weights by ADR 0006. Exact URLs,
hashes, and the distribution decision are in docs/provenance/midas.md.

MIT License

Copyright (c) 2019 Intel ISL (Intel Intelligent Systems Lab)

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.

--------------------------------------------------------------------
Depth Anything V2 (DepthAnything / TikTok)
--------------------------------------------------------------------
Source: https://github.com/DepthAnything/Depth-Anything-V2
License: Apache License 2.0
Copyright (c) 2024 Depth Anything V2 authors.
Citation: Yang, L., Kang, B., Huang, Z., Zhao, Z., Xu, X., Feng, J.,
          and Zhao, H. "Depth Anything V2." NeurIPS 2024.
Used for: Depth Anything V2 model family (DINOv2 encoder + DPT head)
          vendored under libreyolo/models/depth_anything/_vendor/
          (dinov2, dinov2_layers, dpt, util/{blocks,transform}). Bundled
          verbatim except for added package __init__.py files. The
          LibreYOLO-side wrapper (libreyolo/models/depth_anything/
          {model,nn,utils}.py) adds internal ImageNet normalization and
          the depth-task contract; it does not modify the vendored code.

NOTE - code vs. weights: The Apache-2.0 license covers the Depth
Anything V2 *source code* vendored above. It does NOT cover the
pretrained weights, which are split: the Small (ViT-S) checkpoint is
Apache-2.0, while Base/Large/Giant (ViT-B/L/G) are CC-BY-NC-4.0
(non-commercial). LibreYOLO mirrors the converted checkpoints on its
Hugging Face org (LibreYOLO/LibreDepthAnythingV2{s,l}-depth) and
auto-downloads them on demand; each mirror carries the upstream
license. Users remain responsible for complying with each checkpoint's
license — in particular, the CC-BY-NC-4.0 checkpoints (Base/Large/
Giant) are for non-commercial use only. The offline conversion path
(weights/convert_depth_anything_v2_weights.py) remains available.

--------------------------------------------------------------------
Depth Anything 3 (ByteDance Seed)
--------------------------------------------------------------------
Source: https://github.com/ByteDance-Seed/Depth-Anything-3
Pinned commit: 41736238f5bced4debf3f2a12375d2466874866d
License: Apache License 2.0
Copyright (c) 2025 ByteDance Ltd. and/or its affiliates.
Used for: DA3MONO-LARGE model family (ViT-L encoder + DPT depth and sky
          heads) vendored under libreyolo/models/depth_anything3/_vendor/.
          The DINOv2 subcomponents retain Meta Platforms Apache-2.0
          copyright headers. LibreYOLO adaptations remove unused runtime
          dependencies, add internal ImageNet normalization, reproduce the
          official sky handling, and convert positive relative depth to the
          library's relative inverse-depth output contract.

Weight source: https://huggingface.co/depth-anything/DA3MONO-LARGE
Pinned revision: f465978e618db8cc79c83b8bbf24964857db1875
Weight license: Apache License 2.0
Conversion: weights/convert_depth_anything3_weights.py removes only the
            outer model. prefix and wraps 406 unchanged tensors in the
            LibreYOLO checkpoint schema. Only DA3MONO-LARGE is hosted.
            CC-BY-NC-4.0 Large/Giant/Nested weights are excluded.

--------------------------------------------------------------------
ZipDepth (University of Bologna)
--------------------------------------------------------------------
Source: https://github.com/fabiotosi92/ZipDepth
        (commit 6b96f4d205f8a2e5377e81c1b74cc99a47f6693a)
License: MIT
Copyright (c) 2026 Fabio Tosi.
Citation: Tosi, F., Bartolomei, L., Poggi, M., and Mattoccia, S.
          "ZipDepth: Bringing Lightweight Zero-Shot Monocular Depth
          Anywhere, on Any Device." ECCV 2026.
Used for: ZipDepth depth model family (libreyolo/models/zipdepth/).
          The architecture (nn.py) is ported from upstream with
          identical module names; the LibreYOLO wrapper adds the
          depth-task contract, checkpoint schema, zero-shot val, and
          fixed-resolution export metadata.

NOTE - weights: The upstream repository publishes the pretrained
checkpoints (zipdepth_base.pth, zipdepth_base_npu.pth) under the same
MIT license; LibreYOLO mirrors byte-identical rewraps on its Hugging
Face org (LibreYOLO/LibreZipDepth{b,bnpu}-depth). Upstream trained
these weights by distilling pseudo-labels from Depth Anything V2 Large
(itself CC-BY-NC-4.0) over ~14M images from 17 public datasets; the MIT
grant on the student weights is upstream's published position, and the
distillation lineage is documented in
libreyolo/models/zipdepth/NOTICE.

--------------------------------------------------------------------
TorchVision Faster R-CNN (PyTorch contributors)
--------------------------------------------------------------------
Source: https://github.com/pytorch/vision
Commit: 336d36e8db990a905498c73933e35231876e28bc
Version: torchvision 0.26.0
License: BSD-3-Clause
Copyright (c) Soumith Chintala 2016
Used for: the native inference-only Faster R-CNN family in
          libreyolo/models/faster_rcnn/nn.py. The implementation derives its
          two-stage orchestration, RPN, RoI box head, image transform, box
          coder, and FPN-backbone construction from the pinned BSD source.
          The full BSD license text and the code-vs-weights distinction are
          recorded in libreyolo/models/faster_rcnn/NOTICE.

NOTE - code vs. weights: the upstream source code is BSD-3-Clause. The four
official COCO checkpoint files do not carry an explicit checkpoint-specific
license, and torchvision's README warns that pretrained-model terms may differ
based on training data. LibreYOLO's separate weight mirrors apply
BSD-3-Clause on an explicitly disclosed implied basis, ship the verbatim
license with every checkpoint, and repeat the caveat. The source distribution
does not bundle the checkpoint files; conversion only adds metadata.

--------------------------------------------------------------------
TorchVision RetinaNet (PyTorch contributors)
--------------------------------------------------------------------
Source: https://github.com/pytorch/vision
Commit: 336d36e8db990a905498c73933e35231876e28bc
Version: torchvision 0.26.0
License: BSD-3-Clause
Copyright (c) Soumith Chintala 2016
Used for: the native inference-only RetinaNet family in
          libreyolo/models/retinanet/nn.py. The implementation derives its
          one-stage head, P3-P7 anchor generation, box decode, image transform,
          and FPN-backbone construction from the pinned BSD source. The full
          BSD license text and the code-vs-weights distinction are recorded in
          libreyolo/models/retinanet/NOTICE.

NOTE - code vs. weights: the upstream source code is BSD-3-Clause. The two
official COCO checkpoint files do not carry an explicit checkpoint-specific
license, and torchvision's README warns that pretrained-model terms may differ
based on training data. LibreYOLO's separate weight mirrors apply
BSD-3-Clause on an explicitly disclosed implied basis, ship the verbatim
license with every checkpoint, and repeat the caveat. The source distribution
does not bundle the checkpoint files; conversion only adds metadata.

--------------------------------------------------------------------
TorchVision SSD300 (PyTorch contributors)
--------------------------------------------------------------------
Source: https://github.com/pytorch/vision
Commit: 336d36e8db990a905498c73933e35231876e28bc
Version: torchvision 0.26.0
License: BSD-3-Clause
Copyright (c) Soumith Chintala 2016
Used for: the native inference-only SSD family in
          libreyolo/models/ssd/nn.py, preprocessing, default-box decoding,
          class-wise NMS, conversion, validation, and ONNX/backend integration.
          The implementation derives from the pinned SSD300, VGG, anchor,
          box-coder, and transform source. The full BSD license text is in
          libreyolo/models/ssd/NOTICE.

NOTE - code vs. weights: the official COCO checkpoint does not carry an
explicit checkpoint-specific license. LibreYOLO's separate mirror uses
BSD-3-Clause on an explicitly disclosed implied basis and repeats
torchvision's pretrained-model/data warning. Its backbone initialization
traces to Oxford VGG-16 feature weights by Karen Simonyan and Andrew Zisserman,
released under CC BY 4.0. The family notice records the source, license URL,
attribution, and changes. The source distribution does not bundle the
checkpoint; conversion only adds metadata.

--------------------------------------------------------------------
TorchVision Mask R-CNN (PyTorch contributors)
--------------------------------------------------------------------
Source: https://github.com/pytorch/vision
Commit: 336d36e8db990a905498c73933e35231876e28bc
Version: torchvision 0.26.0
License: BSD-3-Clause
Copyright (c) Soumith Chintala 2016
Used for: the native inference-only Mask R-CNN family in
          libreyolo/models/mask_rcnn/nn.py. The mask-specific implementation
          derives its RoIAlign mask head, class-specific mask selection, and
          ResNet-50-FPN v2 construction from the pinned BSD source, on top of
          LibreYOLO's separately attributed native Faster R-CNN graph. The
          full BSD license text and code-vs-weights distinction are recorded
          in libreyolo/models/mask_rcnn/NOTICE.

NOTE - code vs. weights: the upstream source is BSD-3-Clause. The official
COCO checkpoint does not carry an explicit checkpoint-specific license, and
torchvision warns that pretrained-model terms may differ based on training
data. LibreYOLO's separate weight mirror applies BSD-3-Clause on an explicitly
disclosed implied basis, ships the verbatim license, and repeats the caveat.
The source distribution does not bundle the checkpoint file; conversion only
adds metadata.

--------------------------------------------------------------------
TorchVision FCN semantic segmentation (PyTorch contributors)
--------------------------------------------------------------------
Source: https://github.com/pytorch/vision
Commit: 336d36e8db990a905498c73933e35231876e28bc
Version: torchvision 0.26.0
License: BSD-3-Clause
Copyright (c) Soumith Chintala 2016 and the torchvision contributors
Used for: the native inference-only FCN semantic family in
          libreyolo/models/fcn/nn.py. The implementation derives its FCN
          head, graph orchestration, dilated-ResNet builder configuration,
          and feature-layer routing from the pinned BSD source. The full BSD
          license text and the code-vs-weights distinction are recorded in
          libreyolo/models/fcn/NOTICE.

NOTE - code vs. weights: the upstream source code is BSD-3-Clause. The two
official COCO checkpoint files do not carry an explicit checkpoint-specific
license, and torchvision's README warns that pretrained-model terms may differ
based on training data. LibreYOLO's separate weight mirrors apply
BSD-3-Clause on an explicitly disclosed implied basis, ship the verbatim
license with every checkpoint, and repeat the caveat. The source distribution
does not bundle the checkpoint files; conversion only adds metadata.

--------------------------------------------------------------------
CenterNet (Xingyi Zhou)
--------------------------------------------------------------------
Source: https://github.com/xingyizhou/CenterNet
Commit: 4c50fd3a46bdf63dbf2082c5cbb3458d39579e6c
License: MIT, with permissive MIT/BSD-3-Clause lineage
Copyright (c) 2019 Xingyi Zhou
Used for: the native inference-only CenterNet detector in
          libreyolo/models/centernet/nn.py, the fixed-512 BGR affine
          preprocessing in libreyolo/models/centernet/utils.py, and the
          top-100 center decoder in libreyolo/postprocess/centernet.py.

The ResDCN-18 graph follows CenterNet's resnet_dcn.py, which credits
Microsoft's MIT-licensed human-pose-estimation.pytorch. The DLA-34 graph
follows pose_dla_dcn.py and dla.py, whose DLA implementation is
BSD-3-Clause, Copyright (c) 2018 Fisher Yu. The upstream detector also credits
CornerNet (BSD-3-Clause, Copyright (c) 2018 University of Michigan), DCNv2
(BSD-3-Clause, Copyright (c) 2019 Charles Shang), and tf-faster-rcnn (MIT,
Copyright (c) 2017 Xinlei Chen). The complete notices and license texts are
reproduced in libreyolo/models/centernet/NOTICE.

LibreYOLO does not vendor the legacy DCNv2 extension. Native execution uses
torchvision 0.26.0 deform_conv2d from pytorch/vision commit
336d36e8db990a905498c73933e35231876e28bc (BSD-3-Clause, Copyright (c)
Soumith Chintala 2016). The export-only portable implementation was authored
for LibreYOLO and does not copy the extension.

NOTE - code vs. weights: the official COCO checkpoints were published by the
MIT-licensed CenterNet project but have no separate per-object license file.
LibreYOLO's separate mirrors therefore state that MIT is implied by the
releasing project and is not a publisher-confirmed checkpoint-specific grant.
Conversion strips the data-parallel prefix and adds metadata without changing
learned tensors. The source distribution does not bundle checkpoint files.

--------------------------------------------------------------------
TorchVision FCOS (PyTorch contributors)
--------------------------------------------------------------------
Source: https://github.com/pytorch/vision
Commit: 336d36e8db990a905498c73933e35231876e28bc
Version: torchvision 0.26.0
Source file blob: ccbd2496517c33b74a1a1581e0cbf3b3f173bfed
License: BSD-3-Clause
Copyright (c) Soumith Chintala 2016
Used for: the native inference-only FCOS family in
          libreyolo/models/fcos/nn.py, preprocessing in
          libreyolo/models/fcos/utils.py, and postprocessing in
          libreyolo/postprocess/fcos.py. The implementation derives its FCOS
          heads, anchor grid, box decoding, aspect-preserving transform,
          per-level candidate selection, and class-wise NMS from the pinned
          BSD source. The full BSD license text and the code-vs-weights
          distinction are recorded in libreyolo/models/fcos/NOTICE.

NOTE - code vs. weights: the upstream source code is BSD-3-Clause. The
official COCO checkpoint does not carry an explicit checkpoint-specific
license, and torchvision's README warns that pretrained-model terms may differ
based on training data. LibreYOLO's separate weight mirror applies
BSD-3-Clause on an explicitly disclosed implied basis, ships the verbatim
license, and repeats the caveat. The source distribution does not bundle the
checkpoint file; conversion only adds metadata.

--------------------------------------------------------------------
TorchVision AlexNet (PyTorch contributors)
--------------------------------------------------------------------
Source: https://github.com/pytorch/vision
Commit: 336d36e8db990a905498c73933e35231876e28bc
Version: torchvision 0.26.0
License: BSD-3-Clause
Copyright (c) Soumith Chintala 2016
Used for: the native inference-only AlexNet classifier in
          libreyolo/models/alexnet/nn.py. The implementation retains the
          pinned source's features/avgpool/classifier graph and state-dict
          names; LibreYOLO supplies its factory, preprocessing, validation,
          postprocessing, checkpoint metadata, and export surfaces.

NOTE - code vs. weights: the upstream source code is BSD-3-Clause. The
official ImageNet-1K checkpoint does not carry an explicit checkpoint-specific
license. LibreYOLO's separate weight mirror applies BSD-3-Clause on a clearly
disclosed implied basis, includes the verbatim license, and repeats
torchvision's warning that pretrained-model terms may derive from training
data. The source distribution does not bundle the checkpoint; conversion only
adds metadata.

--------------------------------------------------------------------
TorchVision VGG (PyTorch contributors)
--------------------------------------------------------------------
Source: https://github.com/pytorch/vision
Commit: 10f68dbd78b9aa5cab9328f3b2e99cfb0b608122
File: torchvision/models/vgg.py
License: BSD-3-Clause
Copyright (c) Soumith Chintala 2016
Used for: the native inference-only VGG image-classification family in
          libreyolo/models/vgg/nn.py. The implementation derives its feature
          configuration, classifier layout, adaptive pooling, and parameter
          initialization from the pinned BSD source. The full BSD license text
          and the code-vs-weights distinction are recorded in
          libreyolo/models/vgg/NOTICE.

NOTE - code vs. weights: the upstream source code is BSD-3-Clause. The four
official ImageNet-1k V1 checkpoint files do not carry an explicit
checkpoint-specific license, and torchvision warns that pretrained-model terms
can depend on training data. LibreYOLO's separate weight mirrors apply
BSD-3-Clause on an explicitly disclosed implied basis, ship the verbatim
license with every checkpoint, and repeat the caveat. The source distribution
does not bundle the checkpoint files; conversion preserves learned tensors and
only adds metadata.

--------------------------------------------------------------------
TorchVision DeepLabv3 (PyTorch contributors)
--------------------------------------------------------------------
Source: https://github.com/pytorch/vision
Commit: 336d36e8db990a905498c73933e35231876e28bc
Version: torchvision 0.26.0
License: BSD-3-Clause
Copyright (c) Soumith Chintala 2016
Used for: the native inference-only DeepLabv3 family in
          libreyolo/models/deeplabv3/nn.py. The implementation derives its
          segmentation orchestration, ASPP branches, pooling and classifier,
          and ResNet/MobileNetV3 backbone construction from the pinned BSD
          source. The full BSD license text and the code-vs-weights distinction
          are recorded in libreyolo/models/deeplabv3/NOTICE.

NOTE - code vs. weights/data: the upstream source code is BSD-3-Clause. The
three official COCO checkpoint files do not carry an explicit
checkpoint-specific license, and torchvision's README warns that
pretrained-model terms may differ based on training data. LibreYOLO's separate
weight mirrors apply BSD-3-Clause on an explicitly disclosed implied basis,
ship the verbatim license with every checkpoint, and repeat the caveat. COCO
annotations are CC BY 4.0 and its source images retain individual Flickr
terms. The source distribution does not bundle the checkpoint files;
conversion removes only the training-only auxiliary head and adds metadata.

--------------------------------------------------------------------
NAFNet (Megvii Research)
--------------------------------------------------------------------
Source: https://github.com/megvii-research/NAFNet
License: MIT
Copyright (c) 2022 Megvii Inc.
Citation: Chen, L., Chu, X., Zhang, X., and Sun, J. "Simple Baselines
          for Image Restoration." ECCV 2022.
Used for: NAFNet restoration model family (libreyolo/models/nafnet/),
          including the NAFBlock architecture and test-time local
          converter logic. The LibreYOLO wrapper adds the restore-task
          contract, paired train/validation plumbing, fixed-resolution
          ONNX export metadata, and Results.restored payload.

NOTE - code vs. weights/data: The NAFNet source code is MIT licensed.
LibreYOLO does not bundle NAFNet pretrained checkpoint files. Some
published GoPro-trained NAFNet weights do not carry an explicit
standalone weights license; convert only weights that you have the
right to use and redistribute. The GoPro deblurring dataset is separate
from the NAFNet code and carries its own terms; users are responsible
for dataset compliance. The SIDD denoising weights are trained on the
Smartphone Image Denoising Dataset (SIDD), which is distributed under
the MIT License.

--------------------------------------------------------------------
Real-ESRGAN (Xintao Wang)
--------------------------------------------------------------------
Source: https://github.com/xinntao/Real-ESRGAN
License: BSD-3-Clause
Copyright (c) 2021 Xintao Wang.
Citation: Wang, X., Xie, L., Dong, C., and Shan, Y. "Real-ESRGAN:
          Training Real-World Blind Super-Resolution with Pure
          Synthetic Data." ICCV Workshops 2021.
Used for: the Real-ESRGAN super-resolution model family
          (libreyolo/models/realesrgan/). The seam-free tiled forward
          in utils.py is ported from Real-ESRGAN's inference helper.
          The released generator weights (RealESRGAN_x4plus,
          RealESRGAN_x2plus, realesr-general-x4v3) are BSD-3-Clause and
          are mirrored as converted checkpoints with provenance.

--------------------------------------------------------------------
BasicSR (XPixelGroup)
--------------------------------------------------------------------
Source: https://github.com/XPixelGroup/BasicSR
License: Apache License 2.0
Copyright 2018-2022 BasicSR Authors.
Used for: the RRDBNet / SRVGGNetCompact / pixel_unshuffle architecture
          lineage in libreyolo/models/realesrgan/nn.py. Module and
          parameter names mirror BasicSR so the released Real-ESRGAN
          state dicts convert with a plain metadata-wrap.

--------------------------------------------------------------------
SwinIR (Jingyun Liang et al.)
--------------------------------------------------------------------
Source: https://github.com/JingyunLiang/SwinIR
        (commit 6545850fbf8df298df73d81f3e8cba638787c8bd)
License: Apache License 2.0
Copyright 2021 SwinIR Authors.
Citation: Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., and
          Timofte, R. "SwinIR: Image Restoration Using Swin
          Transformer." ICCV Workshops 2021.
Used for: the SwinIR super-resolution generator in
          libreyolo/models/swinir/nn.py. Module and parameter names
          mirror upstream so the official released checkpoints load
          without tensor remapping. The LibreYOLO wrapper adds the
          restore task, native-resolution preprocessing, tiling,
          validation, and checkpoint metadata.

NOTE - weights: the official SwinIR-S x4 lightweight and SwinIR-M/L x4
real-world checkpoints are published with the Apache-2.0 project release.
LibreYOLO does not bundle checkpoint files in the source distribution.

--------------------------------------------------------------------
BiRefNet (Peng Zheng et al.)
--------------------------------------------------------------------
Source: https://github.com/ZhengPeng7/BiRefNet (commit d83f355)
License: MIT
Copyright (c) 2024 ZhengPeng (Peng Zheng).
Citation: Zheng, P., Gao, D., Fan, D.-P., Liu, L., Laaksonen, J.,
          Ouyang, W., and Sebe, N. "Bilateral Reference for
          High-Resolution Dichotomous Image Segmentation." CAAI
          Artificial Intelligence Research, 2024.
Used for: BiRefNet background-removal model family
          (libreyolo/models/birefnet/): the Swin Transformer v1
          backbone and the bilateral-reference decoder (ASPP with
          torchvision deformable convolution). The LibreYOLO port covers
          the inference forward path, adds the matte-task contract
          (Results.matte, cutout, transparent-PNG save), a paired
          MAE/S-measure validator, and fixed-resolution ONNX export.
          Parity verified: our fp32 forward matches the upstream
          released weights with max_abs_diff == 0.

NOTE - code vs. weights: The BiRefNet source code is MIT. The released
BiRefNet (general, Swin-L) weights are tagged MIT on Hugging Face and are
rehosted under the LibreYOLO org. The BiRefNet_lite (Swin-T) Hugging Face
repo shows an MIT badge in its model card but carries no explicit license
metadata (no YAML `license:` field, no LICENSE file); LibreYOLO does not
rehost the lite weights pending an explicit license confirmation. See
weights/LICENSE_NOTICE.txt.

--------------------------------------------------------------------
FeyNobg (Feyn Inc.)
--------------------------------------------------------------------
Source: https://huggingface.co/feyninc/FeyNobg
        (reference library: https://github.com/feyninc/nobg)
License: Apache-2.0 (code and weights)
Copyright (c) 2026 Feyn Inc.
Used for: the FeyNobg background-removal model family
          (libreyolo\models\feynobg\, matte task). FeyNobg is
          architecturally BiRefNet with the third Swin-L stage deepened
          from 18 to 24 blocks; the family reuses the LibreYOLO BiRefNet
          port's nn module with a family-local dimension table and
          converts the released Apache-2.0 weights (state-dict
          metadata-wrap plus post-training-quantized fp8/nvfp4
          variants). No code was copied from the nobg library. Upstream
          training data is not disclosed; LibreYOLO redistributes only
          the released weights.

--------------------------------------------------------------------
PaddleOCR (PaddlePaddle)
--------------------------------------------------------------------
Source: https://github.com/PaddlePaddle/PaddleOCR (commit 211989f)
License: Apache License 2.0
Copyright (c) 2020 PaddlePaddle Authors.
Citation: Cui, C., et al. "PaddleOCR 3.0 Technical Report."
          arXiv:2507.05595, 2025.
Used for: the LibrePPOCR text detection + recognition family
          (libreyolo/models/ppocr/): the PP-LCNetV3 and PP-HGNetV2-B4
          backbones, RSEFPN/LKPAN necks, DB heads, SVTR sequence
          encoder, and CTC head are PyTorch ports of the PP-OCRv5
          Paddle model definitions; the DB quad postprocess and CTC
          greedy decode in libreyolo/postprocess/ppocr.py and the
          det/rec preprocessing in libreyolo/models/ppocr/ follow the
          upstream inference tools. The PP-OCRv5 recognition dictionary
          (ppocr/utils/dict/ppocrv5_dict.txt) is embedded as charset
          metadata in the converted checkpoints. Parity verified: on
          identical input tensors our fp32 forward matches the official
          PP-OCRv5 inference graphs with max_abs_diff <= 1e-4 (det maps)
          and <= 6e-5 (rec probabilities, identical argmax) on both
          tiers.

--------------------------------------------------------------------
DeiT (Facebook Research)
--------------------------------------------------------------------
Source: https://github.com/facebookresearch/deit
Commit: 7e160fe43f0252d17191b71cbb5826254114ea5b
License: Apache License 2.0
Copyright (c) Facebook, Inc. and its affiliates.
Used for: the plain DeiT architecture and released tiny/small/base patch-16
          224-pixel variant contract in libreyolo/models/deit/. The native
          inference code derives from the separately attributed Apache-2.0
          timm implementation below. Distilled, DeiT III, and 384-pixel
          variants are not included.

--------------------------------------------------------------------
Rockchip RKNN Model Zoo
--------------------------------------------------------------------
Source: https://github.com/airockchip/rknn_model_zoo
Commit: bad6c7334531becaf90a561988519b7bec34d0ab
License: Apache License 2.0
Used for: The public RKNN conversion API flow and board-free simulator
          initialization referenced by libreyolo/export/rknn.py. The
          specific upstream reference files were
          examples/LPRNet/python/convert.py and
          py_utils/rknn_executor.py. LibreYOLO's implementation uses its
          own exporter, metadata, calibration, error-handling, and parity
          contracts. No AGPL source code was inspected or used.

NOTE - optional vendor dependency: Rockchip's rknn-toolkit2 compiler and
simulator are not Apache-2.0. They are distributed separately under the
custom RKNN SDK License. LibreYOLO does not bundle, redistribute, install,
or declare that SDK as a project dependency; users review and install it
themselves in a supported Linux x86_64 environment.

--------------------------------------------------------------------
timm / PyTorch Image Models (Hugging Face)
--------------------------------------------------------------------
Source: https://github.com/huggingface/pytorch-image-models
License: Apache License 2.0
Copyright (c) Ross Wightman and the timm contributors.
Used for: the native image-classification model families ported from
          timm architectures — libreyolo/models/{mobilenetv4,convnext,deit,
          efficientnetv2,resnet,swin,vit}/ — the MiDaS ViT-L/16 and
          EfficientNet-Lite3 encoders at libreyolo/models/midas/, and the
          shared Swin backbone at libreyolo/models/swin/. Module/attribute
          names mirror timm so
          its Apache-2.0 ImageNet-1k pretrained weights load unchanged and
          inference is bit-identical. The unmodified Apache-2.0 timm
          data/_info/imagenet_synsets.txt index is included as
          libreyolo/data/imagenet1k_synsets.txt so standard WNID ImageFolder
          datasets map to those classifier indices. Architecture lineage:
          ConvNeXt also derives from facebookresearch/ConvNeXt (MIT); DeiT
          from facebookresearch/deit (Apache-2.0); EfficientNetV2 from
          google/automl (Apache-2.0); ResNet from He et al. 2015. Weights
          (timm *.in1k / *.fb_in1k / *.augreg*_in21k_ft_in1k / a1_in1k,
          Apache-2.0) are mirrored on
          the LibreYOLO Hugging Face org. ConvNeXt-V2 fcmae weights
          (CC-BY-NC) are NOT used.

          LibreViT specifically derives its native fixed-224 patch-16 graph
          from timm v1.0.28, commit
          8ef73809f622e0031bd7f4940265734aef8b9978. Its AugReg checkpoint
          revisions and source/converted SHA-256 hashes are recorded in
          docs/provenance/vit.md.

Swin Transformer V1 model and weight lineage: microsoft/Swin-Transformer at
commit f82860bfb5225915aca09c3227159ee9e1df874d and
SwinTransformer/storage release v1.0.0 at commit
3cc359915d3a6079b176a871f68d5fb0d8dfdea2, both MIT. The Tiny, Small, Base, and
Large patch-4/window-7 learned tensors are mirrored through timm's pinned
Hugging Face repositories and converted without numerical changes. Large was
pretrained on ImageNet-22k and fine-tuned on ImageNet-1k. See
libreyolo/models/swin/NOTICE for exact source revisions.

--------------------------------------------------------------------
EfficientDet (rwightman/efficientdet-pytorch; Google Research)
--------------------------------------------------------------------
Source: https://github.com/rwightman/efficientdet-pytorch
Commit: c6dff775a36cea0bf9b76c58e59f936411c5ce01
Version: effdet 0.4.1
Backbone source: https://github.com/huggingface/pytorch-image-models
Backbone version/commit: v1.0.28 / 8ef73809f622e0031bd7f4940265734aef8b9978
Original source: https://github.com/google/automl/tree/master/efficientdet
License: Apache License 2.0
Copyright (c) 2020 Ross Wightman; Copyright 2020 Google Research.
Used for: the EfficientDet D0-D4 inference family
          (libreyolo/models/efficientdet/), centralized anchor decode and NMS
          (libreyolo/postprocess/efficientdet.py), evaluation preprocessing,
          strict checkpoint converter, and one-output export wrapper. Official
          D0-D4 COCO weights are metadata-wrapped without changing learned
          tensors. The release assets have no separate weight license; their
          redistribution basis is Apache-2.0 implied by the releasing project.
          No source from the LGPL-licensed
          zylo117/Yet-Another-EfficientDet-Pytorch project was consulted or
          used.

--------------------------------------------------------------------
Apache License 2.0 (full text)
--------------------------------------------------------------------
The full text of the Apache License, Version 2.0 is bundled with this
distribution at licenses/Apache-2.0.txt (also available at
https://www.apache.org/licenses/LICENSE-2.0) and applies to the
Apache-2.0 upstreams listed above.

--------------------------------------------------------------------
L2CS-Net
--------------------------------------------------------------------
Source: https://github.com/Ahmednull/L2CS-Net
License: MIT
Copyright (c) 2022 Ahmed Abdelrahman
Citation: Abdelrahman, A. A., Hempel, T., Khalifa, A., Al-Hamadi, A.,
          and Dinges, L. "L2CS-Net: Fine-Grained Gaze Estimation in
          Unconstrained Environments." IEEE International Conference
          on Image Processing (ICIP), 2022.
Used for: L2CS gaze estimation network (libreyolo/models/l2cs/nn.py),
          bin-expectation angle decoding and crop preprocessing
          (libreyolo/models/l2cs/utils.py), and gaze arrow visualization
          (libreyolo/utils/drawing.py:draw_gaze_arrows).

NOTE — code vs. weights: The MIT license below covers the L2CS-Net
*source code*, which is what libreyolo/models/l2cs/ is ported from.
It does NOT cover the pretrained weights. The published L2CS gaze
checkpoints (e.g. L2CSNet_gaze360.pkl) are trained on the Gaze360
dataset and are bound by the Gaze360 dataset license — research /
non-commercial use only, no redistribution:
  https://github.com/erkil1452/gaze360/blob/master/LICENSE.md
LibreYOLO therefore does NOT bundle, mirror, or auto-download L2CS
weights. Users obtain them from the official L2CS-Net distribution and
are responsible for complying with the Gaze360 license. Required
dataset citation: Kellnhofer, Recasens, Stent, Matusik, Torralba,
"Gaze360: Physically Unconstrained Gaze Estimation in the Wild",
ICCV 2019.

MIT License

Copyright (c) 2022 Ahmed Abdelrahman

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.

--------------------------------------------------------------------
Darknet (YOLOv1 / YOLOv2 / YOLOv3 / YOLOv4)
--------------------------------------------------------------------
Source: https://github.com/pjreddie/darknet   (YOLOv1/v2/v3)
        https://github.com/AlexeyAB/darknet    (YOLOv4)
License: Public domain ("YOLO LICENSE")
Used for: the LibreYOLO1 / LibreYOLO2 / LibreYOLO3 / LibreYOLO4 families
          (libreyolo/models/darknet, libreyolo/models/yolo{1,2,3,4}).
          The public-domain .cfg model definitions are bundled under
          libreyolo/models/darknet/cfgs/; only the .cfg format and the
          numerical behaviour of the Darknet layers are reproduced. No
          Darknet C source is copied.

YOLO LICENSE, Version 2, July 29 2016

    0. Darknet is public domain.
    1. Do whatever you want with it.
    2. Stop emailing me about it!

--------------------------------------------------------------------
MultimediaTechLab/YOLO (YOLOv7, YOLOv9)
--------------------------------------------------------------------
Source: https://github.com/MultimediaTechLab/YOLO
License: MIT
Copyright (c) 2024 Kin-Yiu, Wong and Hao-Tang, Tsui
Used for: the LibreYOLO7 family (native port of YOLOv7,
          libreyolo/models/yolo7) and the LibreYOLO9 detection head. This is
          the original authors' MIT re-release, NOT the GPL-3.0
          WongKinYiu/yolov7 or WongKinYiu/yolov9. Module names mirror upstream
          so the MIT-licensed weights load unchanged; the v7.yaml model
          definition is bundled under libreyolo/models/yolo7/.

MIT License

Copyright (c) 2024 Kin-Yiu, Wong and Hao-Tang, Tsui

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, subject to the inclusion of the above
copyright notice and this permission notice. THE SOFTWARE IS PROVIDED "AS IS",
WITHOUT WARRANTY OF ANY KIND.

--------------------------------------------------------------------
Roboflow Trackers (BoT-SORT)
--------------------------------------------------------------------
Source: https://github.com/roboflow/trackers
Commit: 3b6d910df78a7ab48d5770b5bc86e043476d2e76
License: Apache License 2.0
Copyright (c) 2026 Roboflow. All Rights Reserved.
Used for: the BoT-SORT tracking lifecycle, scale-aware center-width-height
          Kalman model, and sparse optical-flow camera-motion compensation in
          libreyolo/tracking/botsort.py and
          libreyolo/tracking/kalman_filter.py. The implementation is adapted
          to LibreYOLO's detector-agnostic Results contract and implements the
          paper's motion-only BoT-SORT variant (not BoT-SORT-ReID).

The full Apache License 2.0 text is included at licenses/Apache-2.0.txt.

--------------------------------------------------------------------
Torchreid (deep-person-reid)
--------------------------------------------------------------------
Source: https://github.com/KaiyangZhou/deep-person-reid
Commit: f8cd150fdf77e8d9e1ed143b7f308c2c609ded50
License: MIT
Copyright (c) 2018 Kaiyang Zhou
Used for: the OSNet-AIN appearance (ReID) embedder used by the Deep
          OC-SORT tracker (libreyolo/tracking/reid.py). Module names
          mirror upstream torchreid/models/osnet_ain.py so the released
          checkpoints load unchanged (bit-exact forward parity, see
          tests/unit/test_reid.py).

MIT License

Copyright (c) 2018 Kaiyang Zhou

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, subject to the inclusion of the above
copyright notice and this permission notice. THE SOFTWARE IS PROVIDED "AS IS",
WITHOUT WARRANTY OF ANY KIND.

--------------------------------------------------------------------
Deep OC-SORT
--------------------------------------------------------------------
Source: https://github.com/GerardMaggiolino/Deep-OC-SORT
Commit: 6bb51d027b137233f5c520b6fcc4f2ae387a6ba9
License: MIT
Copyright (c) 2023 Gerard Maggiolino
Used for: the adaptive appearance association (dynamic embedding EMA and
          adaptive weighting) in the Deep OC-SORT tracker
          (libreyolo/tracking/deepocsort.py), validated for numeric
          track-ID parity against upstream
          (tests/unit/test_deepocsort_parity.py).

MIT License

Copyright (c) 2023 Gerard Maggiolino

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, subject to the inclusion of the above
copyright notice and this permission notice. THE SOFTWARE IS PROVIDED "AS IS",
WITHOUT WARRANTY OF ANY KIND.

--------------------------------------------------------------------
MoGe-2 (Microsoft)
--------------------------------------------------------------------
Source: https://github.com/microsoft/MoGe
Commit: 925b8ed835a7a9cdb7578ba15c658a0afc969030
License: MIT
Copyright (c) Microsoft Corporation.
Used for: the MoGe-2 ViT-S/B/L surface-normal encoder wrapper, image-plane
          coordinate construction, convolutional neck, normal decoder, and
          normal-only inference graph in libreyolo/models/moge2/. The point,
          mask, and metric-scale heads are omitted. The DINOv2 backbone reuses
          LibreYOLO's existing Apache-2.0 implementation rather than copying
          another upstream implementation.

MIT License

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.

--------------------------------------------------------------------
PicoSAM3
--------------------------------------------------------------------
Source: https://github.com/pbonazzi/picosam3
Commit: 1b03949e43472953bb0021685c7fc3f5fdf48fde
License: Apache License 2.0
Used for: the native LibrePicoSAM3 ROI segmentation network in
          libreyolo/models/picosam3. The port uses the upstream
          depthwise-separable encoder-decoder, dilated bottleneck, ECA block,
          ROI preprocessing geometry, and ImageNet normalization. It does not
          vendor the repository's SAM teacher implementations or cctorch code.

The full Apache License 2.0 text is included at
libreyolo/models/picosam3/LICENSE.

--------------------------------------------------------------------
EdgeTAM / Hugging Face Transformers EdgeTAM converter
--------------------------------------------------------------------
Model source: https://github.com/facebookresearch/EdgeTAM
Model commit: 7711e012a30a2402c4eaab637bdb00a521302c91
Converter source: https://github.com/huggingface/transformers
Converter commit: bd37c453544e83eb875ed3608980a1660376007a
Converter file:
  src/transformers/models/edgetam_video/convert_edgetam_video_to_hf.py
Payload reference: https://huggingface.co/yonigozlan/EdgeTAM-hf
Payload revision: c266ce53b3fc00f0f495b583f6a116c4e57f53bb
License: Apache License 2.0
Copyright (c) Meta Platforms, Inc. and affiliates
Copyright 2025 The Hugging Face Inc. team
Used for: the Transformers-backed LibreEdgeTAM adapter in
          libreyolo/models/sam/edgetam.py and the lossless checkpoint mapping
          in weights/convert_edgetam_weights.py. No EdgeTAM architecture source
          is vendored. The runtime adapter reproduces the pinned square image
          transform and prompt-coordinate scaling from sam2/utils/transforms.py.
          The conversion remaps keys and splits/concatenates tensors so the
          official Apache-2.0 checkpoint can be loaded by Transformers.
          Hash-pinned configuration and processor files are copied from the
          Apache-2.0-declared payload reference; learned tensors are converted
          independently and then checked exactly against it.

The full Apache License 2.0 text is included at licenses/Apache-2.0.txt.

--------------------------------------------------------------------
SenseNova-Vision / Bagel (SenseTime; ByteDance; Hugging Face; BFL)
--------------------------------------------------------------------
Source: https://github.com/OpenSenseNova/SenseNova-Vision
        (commit 12ccd96e32b32967a11cacb6c5bd5fe3a555fc0c)
License: Apache License 2.0 (see licenses/Apache-2.0.txt)
Copyright (c) 2026 SenseTime Group Inc.; (c) 2025 Bytedance Ltd.;
          (c) 2024 The Qwen Team and The HuggingFace Inc. team;
          (c) 2024 Black Forest Labs
Citation: SenseNova-Vision team. "Vision as Unified Multimodal
          Generation." arXiv:2607.06560, 2026.
Used for: the LibreSenseNovaVision family (libreyolo/models/sensenova/), an
inference-only port of the Bagel-MoT unified multimodal architecture, its
interleaved inferencer, image transforms, task prompts, and structured-output
parsers. The upstream file modeling/bagel/modeling_utils.py is CC BY-NC 4.0
(DiT-derived) and is NOT ported; its standard components are re-derived from
Hugging Face transformers (ViT-MAE sincos table, Apache-2.0) and
openai/guided-diffusion (timestep embedding, MIT) in modeling/layers.py.
Model weights (sensenova/SenseNova-Vision-7B-MoT, CC BY-NC 4.0,
non-commercial) are mirrored byte-identically with attribution at
huggingface.co/LibreYOLO/SenseNovaVision7b; mirroring does not change the
license. See libreyolo/models/sensenova/NOTICE for the full provenance chain.

--------------------------------------------------------------------
MODUS (EPFL VILAB)
--------------------------------------------------------------------
Source: https://github.com/EPFL-VILAB/Modus
Commit: c299ef0fbba1cfe7c93336c45d7085afd770c0fa
License: Apache License 2.0 (code; see licenses/Apache-2.0.txt)
Copyright 2026 EPFL Visual Intelligence and Learning Lab (VILAB) and the
MODUS authors; Copyright 2025 Bytedance Ltd. and/or its affiliates;
Copyright The HuggingFace Inc. team and the Qwen team.
Used for: the LibreMODUS analysis-only any-to-any family
          (libreyolo/models/modus/): the 16-modality registry and tokenizer
          order, context/sampler logic, and constrained detection decoders are
          adapted from the pinned Apache-2.0 source.
          The shared Bagel/Qwen2/SigLIP/FLUX runtime comes from the audited
          LibreSenseNovaVision port described above.

The upstream CC BY-NC 4.0 modeling/bagel/modeling_utils.py file is not used.
Its standard deterministic utilities are clean permissive implementations
whose sources are documented in libreyolo/models/sensenova/NOTICE.

Model weights are not covered by this code-license declaration. LibreYOLO does
not bundle, mirror, rename, or publish them. It loads local files or downloads
directly from EPFL-VILAB/MODUS at Hugging Face revision
8428a81602c19141e422b1e1795dddcb5d2bc14b. The upstream model card declares
license ``other``, calls the checkpoint ``bagel-derived``, and requests
research-only use. Local FP8 caches remain on the user's machine and retain
the source checkpoint's terms. Fresh downloads require credentials from the
user's own Hugging Face account. See libreyolo/models/modus/NOTICE.

--------------------------------------------------------------------
AuraFace-v1 (fal.ai)
--------------------------------------------------------------------
Source: https://huggingface.co/fal/AuraFace-v1
License: Apache-2.0 (weights)
Used for: the librefacerec-l face-embedding ONNX weights (single file
glintr100.onnx, mirrored unmodified to the LibreYOLO Hugging Face org
as librefacerec-l.onnx). No code is ported; the graph is consumed
opaquely via onnxruntime. Upstream training data is undisclosed; only
this single Apache-2.0 file is mirrored, none of the other artifacts
in the upstream repository.

--------------------------------------------------------------------
YuNet face detector (OpenCV Zoo)
--------------------------------------------------------------------
Source: https://github.com/opencv/opencv_zoo/tree/main/models/face_detection_yunet
License: MIT (weights), Copyright (c) Shiqi Yu
Used for: the librefacerec-det default face-detector ONNX weights
(face_detection_yunet_2023mar.onnx, mirrored unmodified to the
LibreYOLO Hugging Face org as librefacerec-det.onnx). Consumed via
OpenCV's bundled cv2.FaceDetectorYN; no code is ported.
MHR / Momentum Human Rig (Meta Platforms, Inc.)
--------------------------------------------------------------------
Source: https://github.com/facebookresearch/MHR
License: Apache License 2.0 (see licenses/Apache-2.0.txt)
Copyright (c) Meta Platforms, Inc. and affiliates.
Used for: the body model behind the `mesh` task. No MHR code is ported. The
TorchScript asset assets/mhr_model.pt is downloaded at runtime from the public
upstream release by libreyolo/models/sam3dbody/mhr_body.py and cached locally;
it is NOT mirrored by LibreYOLO. The parameter layout used to drive it (204
model parameters, 45 identity coefficients, 72 expression coefficients) is
documented by the upstream Apache-2.0 demo.py and mhr.py and was additionally
verified empirically against the released asset.

--------------------------------------------------------------------
SAM 3D Body (Meta Platforms, Inc.)
--------------------------------------------------------------------
Source: https://github.com/facebookresearch/sam-3d-body
License: SAM License (NOT a permissive license; it carries field-of-use
         restrictions including military, nuclear, espionage and weapons use,
         a no-reverse-engineering clause, and unilateral amendment by Meta)
Copyright (c) Meta Platforms, Inc. and affiliates.
Used for: NO CODE FROM THIS PROJECT IS PORTED, ADAPTED OR VENDORED. Because
the SAM License is not permissive, LibreYOLO deliberately wraps rather than
ports it. libreyolo/models/sam3dbody/ is LibreYOLO's own MIT code that calls
the upstream package's public API and maps its output into the Meshes result
payload. The upstream package is an OPTIONAL dependency the user installs
themselves; LibreYOLO ships none of it, and users who do not use the mesh task
never encounter its terms.
Model weights: the SAM License permits redistribution provided its terms are
passed through, so the checkpoints are mirrored byte-identically at
huggingface.co/LibreYOLO/LibreSAM3DBodyd3-mesh and
huggingface.co/LibreYOLO/LibreSAM3DBodyh-mesh, each carrying the SAM License
text and an access gate recording acceptance. Mirroring does not change the
license: those weights remain SAM-licensed, not MIT.

--------------------------------------------------------------------
TEED (Xavier Soria Poma)
--------------------------------------------------------------------
Source: https://github.com/xavysp/TEED
Commit: 40fa4b1391dc6424f88989d0ca75d5b592c8681d
License: MIT
Copyright (c) 2022 Xavier Soria Poma
Used for: the native TEED edge-detection architecture in
          libreyolo/models/teed/nn.py. The port retains the released
          parameter layout for strict local checkpoint conversion. The full
          MIT notice and the source/weights/data separation are recorded in
          libreyolo/models/teed/NOTICE.

--------------------------------------------------------------------
DexiNed (Xavier Soria Poma)
--------------------------------------------------------------------
Source: https://github.com/xavysp/DexiNed
Commit: 08ed67ad0579f3969536a9719cdc1b829fb74fc1
License: MIT
Copyright (c) 2019 Xavier Soria Poma
Used for: the native DexiNed edge-detection architecture in
          libreyolo/models/dexined/nn.py. The port retains the released
          parameter layout for strict local checkpoint conversion. The full
          MIT notice and the source/weights/data separation are recorded in
          libreyolo/models/dexined/NOTICE.

--------------------------------------------------------------------
HRNet pose (Leo Xiao / Microsoft)
--------------------------------------------------------------------
Source: https://github.com/leoxiaobin/deep-high-resolution-net.pytorch
Commit: 6f69e4676ad8d43d0d61b64b1b9726f0c369e7b1
License: MIT
Repository license: Copyright (c) 2019 Leo Xiao
Adapted source headers: Copyright (c) Microsoft. Written by Bin Xiao.
Used for: the native HRNet W32/W48 COCO-17 pose graph, affine person-crop
          geometry, heatmap decoding, flip restoration, and OKS suppression
          in libreyolo/models/hrnet/ and libreyolo/postprocess/hrnet.py.
          Exact file mapping and hashes are in docs/provenance/hrnet.md.

MIT License

Copyright (c) 2019 Leo Xiao

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
