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.

ConvNeXt V2 native classifier

Adapted from facebookresearch/ConvNeXt-V2
https://github.com/facebookresearch/ConvNeXt-V2
Commit: 2553895753323c6fe0b2bf390683f5ea358a42b9
Files: models/convnextv2.py; dense LayerNorm and GRN from models/utils.py.
Code license: MIT. Official ImageNet pretrained/fine-tuned weights:
CC-BY-NC-4.0, redistributed separately on Hugging Face with the full license.

Modifications: removed sparse pretraining and timm runtime dependencies;
added native family registration, classification head reset, conversion,
preprocessing reuse, and supervised fine-tuning integration. Dense forward
arithmetic and official tensor names are preserved. The supervised recipe
uses zero stochastic depth; no FCMAE pretraining implementation is included.

MIT License
=======================================================================

Copyright (c) Meta Platforms, Inc. and affiliates.

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.
=======================================================================

--------------------------------------------------------------------
LeVJEPA architecture and Hugging Face Transformers
--------------------------------------------------------------------
Architecture paper: https://arxiv.org/abs/2608.27395
Upstream metadata: https://github.com/MLO-lab/LeVJEPA
Commit: 941526c428aa513a5bdfa38697724fda794d8496
Permissive implementation source: https://github.com/huggingface/transformers
Commit: 3fa4da70f5d1da3ab1a8304bd2339eab7004b157
License: Apache-2.0
Used for: standard transformer and factorized rotary primitives in
          libreyolo/models/levjepa/, through LibreYOLO's existing V-JEPA 2
          implementation.

The LeVJEPA inference graph is independently implemented from the paper and
factual checkpoint metadata. No code from upstream's CC BY-NC 4.0 module.py or
Hugging Face remote-code implementation was read, copied, adapted, or bundled.
The learned weights are distributed separately under CC BY-NC 4.0; see
weights/LICENSE_NOTICE.txt and the family NOTICE.
The full Apache License 2.0 text is included at licenses/Apache-2.0.txt.

--------------------------------------------------------------------
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.

--------------------------------------------------------------------
DDColor and permissive architecture lineage
--------------------------------------------------------------------
Primary source: https://github.com/piddnad/DDColor
Commit: 2adb63f2656ac41cbdf7b894cddd94121a3faf13
License: Apache License 2.0
Used for: the inference-only ConvNeXt encoder, wide UNet decoder,
          multi-scale color transformer, Lab preprocessing/postprocessing,
          and compatible parameter layout in libreyolo/models/ddcolor/.

Derived component sources:
- BasicSR, https://github.com/XPixelGroup/BasicSR, commit
  8d56e3a045f9fb3e1d8872f92ee4a4f07f886b0a, Apache-2.0,
  Copyright 2018-2022 BasicSR Authors.
- ColorFormer, https://github.com/jixiaozhong/ColorFormer, commit
  0d3ae850a3990bbdb954bb39feba07b15c1aa357, permissive three-clause license,
  Copyright (C) 2022 THL A29 Limited, a Tencent company.
- fastai, https://github.com/fastai/fastai, commit
  517e68da4779897e6d2668f07aa9aed52e0e4a89, Apache-2.0.
- ConvNeXt, https://github.com/facebookresearch/ConvNeXt, commit
  048efcea897d999aed302f2639b6270aedf8d4c8, MIT,
  Copyright (c) Meta Platforms, Inc. and affiliates.
- Mask2Former, https://github.com/facebookresearch/Mask2Former, commit
  9b0651c6c1d5b3af2e6da0589b719c514ec0d69a, MIT,
  Copyright (c) 2022 Meta, Inc.
- DETR, https://github.com/facebookresearch/detr, commit
  29901c51d7fe8712168b8d0d64351170bc0f83e0, Apache-2.0,
  Copyright (c) Facebook, Inc. and its affiliates.
- BigColor, https://github.com/KIMGEONUNG/BigColor, commit
  5b7687fbedd9af22fef02a82102c6d7708ab1217, MIT, acknowledged architectural
  and training lineage only; no BigColor source is copied.

The exact adapted paths are recorded in libreyolo/models/ddcolor/NOTICE.
Training code, private data, and the Artistic checkpoint are excluded.

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

ColorFormer three-clause license
--------------------------------

Copyright (C) 2022 THL A29 Limited, a Tencent company. 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.

3. Neither the name of the copyright holder nor the names of its contributors
   may be used to endorse or promote products derived from this software
   without specific prior written permission.

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 HOLDER 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.

MIT component notices for ConvNeXt and Mask2Former
--------------------------------------------------

Copyright (c) Meta Platforms, Inc. and affiliates.
Copyright (c) 2022 Meta, Inc.

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.

--------------------------------------------------------------------
HVI-CIDNet
--------------------------------------------------------------------
Source: https://github.com/Fediory/HVI-CIDNet
Commit: eb43d7d91e9a336c66856824ff9e4603ae41f408
License: MIT
Copyright (c) 2024 Fediory Feng
Used for: the CIDNet inference graph, HVI transform, local-context attention,
          and compatible parameter layout in libreyolo/models/hvi_cidnet/.
          Training, loss, metric, and dataset code are not included.

The full source and checkpoint/data separation is recorded in
libreyolo/models/hvi_cidnet/NOTICE. LibreYOLO's own wrapper, validation,
conversion, and tests remain under the repository's MIT license.

Permissive architecture lineage:
- Restormer, commit 68dc6ac472db26f16361150cb7a96a1bc87da93f,
  MIT, Copyright (c) 2022 Syed Waqas Zamir and contributors.
- ConvNeXt, commit 048efcea897d999aed302f2639b6270aedf8d4c8,
  MIT, Copyright (c) Meta Platforms, Inc. and affiliates.
- einops, commit e0d5eb4fd535945ff65d309206d0f1754a926821,
  MIT, Copyright (c) 2018 Alex Rogozhnikov.

The primary repository's eval_hf.py is behaviorally referenced for
preprocessing and crop semantics; the LibreYOLO helper is a native
implementation. Each lineage project uses the same MIT grant reproduced
below for the adapted HVI-CIDNet source.

MIT License

Copyright (c) 2024 Fediory Feng

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.

--------------------------------------------------------------------
LaMa OpenCV Zoo ONNX artifact
--------------------------------------------------------------------
Artifact source: https://github.com/opencv/opencv_zoo
Commit: 966be4cb6eefc2a4f586ed0581bf4bbaa2533070
Path: models/inpainting_lama
License: Apache License 2.0
Directory notice: Copyright [2021] Samsung Research
Conversion lineage: https://huggingface.co/Carve/LaMa-ONNX
Revision: c3c0c9e468934d62e79c329e35d82dd09ff8c444
Original method: https://github.com/advimman/lama
Commit: 786f5936b27fb3dacd2b1ad799e4de968ea697e7
Used for: the immutable QDQ ONNX graph embedded in LibreLaMab-restore.pt.

No LaMa source code is copied or reimplemented. The independently written
MIT adapter in libreyolo/models/lama/ executes the artifact through MIT-
licensed ONNX Runtime. Exact artifact and training-data provenance is in
libreyolo/models/lama/NOTICE. The full Apache License 2.0 text is included at
licenses/Apache-2.0.txt.

--------------------------------------------------------------------
ViTMatte / Hugging Face Transformers
--------------------------------------------------------------------
Port source: https://github.com/huggingface/transformers
Commit: 7d6354e04794f3246bf9a0faf4fead080edeebb6
License: Apache License 2.0
Copyright 2023 HUST-VL and The HuggingFace Inc. team
Copyright 2023 The HuggingFace Inc. team
Copyright 2023 Meta AI and The HuggingFace Inc. team
Architecture lineage: https://github.com/hustvl/ViTMatte
Commit: f48208ceb0d172ab0c1c528bba33cc1159596407
License: MIT
Copyright (c) 2023 Hust Vision Lab
Used for: the ViTMatte-S/ViTDet inference graph and four-channel image/trimap
          processing contract in libreyolo/models/vitmatte/.

No HUST training-stack code or its optional dependencies are copied. The
Composition-1k pretrained checkpoint has separate Adobe non-commercial terms;
see libreyolo/models/vitmatte/NOTICE and weights/LICENSE_NOTICE.txt. The full
Apache License 2.0 text is included at licenses/Apache-2.0.txt.

--------------------------------------------------------------------
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.

--------------------------------------------------------------------
Perception Encoder / PE Core (Meta AI)
--------------------------------------------------------------------
Source: https://github.com/huggingface/pytorch-image-models (tag v1.0.28,
        commit 8ef73809f622e0031bd7f4940265734aef8b9978)
        https://github.com/mlfoundations/open_clip (tag v3.2.0,
        commit 6f939057c792a2f3d4d58df748de60ca47c4aed4)
License: Apache License 2.0 (pytorch-image-models); MIT License (open_clip)
Copyright (c) Ross Wightman; (c) 2012-2021 OpenCLIP authors.
Used for: the LibrePE family (libreyolo/models/pe/). The vision tower
(PEVisionTransformer, PEBlock, PEAttentionRope, PEAttentionPoolLatent,
RotaryEmbeddingCat and the rotary/fourier helpers in libreyolo/models/pe/nn.py)
is adapted from timm's EVA implementation (timm/models/eva.py,
timm/layers/pos_embed_sincos.py, timm/layers/attention_pool.py). The text tower
(PETextTransformer, Transformer, ResidualAttentionBlock in the same file) is
adapted from open_clip's src/open_clip/transformer.py. Neither timm nor
open_clip is imported at runtime. The OpenAI BPE tokenizer is reused from the
existing libreyolo/models/clip/ implementation rather than duplicated.

The Perception Encoder architecture itself is described in
"Perception Encoder: The best visual embeddings are not at the output of the
network" (Meta AI, https://arxiv.org/abs/2504.13181).

NOTE: the official facebookresearch/perception_models repository (commit
3e352cca660658d4b5c90f42a7808b11469e4c66) ships LICENSE.PE with Apache-2.0
terms while the same revision's setup.py declares a noncommercial/proprietary
package license. That inconsistency is unresolved, and is why no code was
copied or adapted from that repository. The Perception Language Model
(LICENSE.PLM, noncommercial), PE Spatial and PE-AV are out of scope.

The shipped LibrePE weights are converted from the Apache-2.0
OpenCLIP-compatible timm/PE-Core-* repositories (state-dict metadata wrap only;
learned parameters unchanged). See libreyolo/models/pe/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.

--------------------------------------------------------------------
mmsegmentation UNet (OpenMMLab)
--------------------------------------------------------------------
Source: https://github.com/open-mmlab/mmsegmentation
        pinned commit b040e147adfa
        (mmseg/models/backbones/unet.py,
         mmseg/models/utils/up_conv_block.py,
         mmseg/models/decode_heads/fcn_head.py)
License: Apache License 2.0
Copyright (c) OpenMMLab. All rights reserved.
Used for: U-Net semantic segmentation family
          (libreyolo/models/unet/). Native same-padded UNet-S5-D16 +
          FCN-head graph whose state-dict keys match the official
          Cityscapes checkpoint. This is not the 2015 Caffe
          valid-convolution U-Net and not nnU-Net. The released
          Cityscapes checkpoint keeps the Apache-2.0 license
          mmsegmentation declares for it; see
          weights/LICENSE_NOTICE.txt and libreyolo/models/unet/NOTICE.

--------------------------------------------------------------------
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 / PP-LiteSeg
--------------------------------------------------------------------
Source: https://github.com/Deci-AI/super-gradients
        pinned commit 63de22c404d5740f34f7706c302b37fce3c8fe5d
License: Apache License 2.0
Copyright (c) 2021-2024 Deci AI
Used for: PP-LiteSeg semantic segmentation family
          (libreyolo/models/ppliteseg/): the STDC backbone, SPPM context
          module, UAFM decoder, segmentation and auxiliary heads, the
          Dice + cross-entropy + edge compound loss, and the Cityscapes
          training recipe. Architecture cross-checked against
          PaddlePaddle/PaddleSeg (Apache-2.0, commit 3c4db66d); no
          PaddleSeg code copied. The released PP-LiteSeg checkpoints are
          trained on Cityscapes and are redistributed only under the
          Cityscapes non-commercial terms, which do not apply to the
          Apache-2.0 code above; see weights/LICENSE_NOTICE.txt and
          libreyolo/models/ppliteseg/NOTICE.

--------------------------------------------------------------------
SuperGradients / training safeguards and helpers
--------------------------------------------------------------------
Source: https://github.com/Deci-AI/super-gradients
        pinned commit 63de22c404d5740f34f7706c302b37fce3c8fe5d
License: Apache License 2.0 (repository LICENSE.md at that commit)
Copyright (c) 2021-2024 Deci AI
Used for: original LibreYOLO implementations of the QAT-time EMA and
          SyncBatchNorm disable policy, metric-gated top-N checkpoint
          averaging, precise BatchNorm re-estimation, repeat-factor
          class-balanced sampling, and an epoch-0 ONNX export check. Code
          lives in libreyolo/training/qat_defaults.py,
          libreyolo/training/weight_averaging.py,
          libreyolo/training/precise_bn.py,
          libreyolo/training/export_check.py, and
          libreyolo/data/class_balanced.py. No upstream source code was
          copied.

--------------------------------------------------------------------
STDC-Seg (MichaelFan01)
--------------------------------------------------------------------
Source: https://github.com/MichaelFan01/STDC-Seg
        pinned commit 59ff37fbd693b99972c76fcefe97caa14aeb619f
License: MIT
Copyright (c) 2021 Mingyuan Fan
Used for: STDC backbone lineage in the PP-LiteSeg family
          (libreyolo/models/ppliteseg/nn.py). The adapted SuperGradients
          stdc.py states it is based on this exact commit, so the MIT
          notice travels with the STDC-derived blocks.

--------------------------------------------------------------------
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.

--------------------------------------------------------------------
SuperGradients / DEKR-W32-NO-DC (bottom-up pose)
--------------------------------------------------------------------
Source: https://github.com/Deci-AI/super-gradients
        pinned commit 63de22c404d5740f34f7706c302b37fce3c8fe5d
License: Apache License 2.0 (repository LICENSE.md at that commit)
Copyright (c) 2021-2024 Deci AI
Additional notice: the DEKR model file at that pin carries its own
        upstream header, reproduced here and preserved in
        libreyolo/models/dekr/nn.py:
            Copyright (c) Microsoft
            Licensed under the MIT License.
            The code is based on HigherHRNet-Human-Pose-Estimation.
            (https://github.com/HRNet/HigherHRNet-Human-Pose-Estimation)
            Modified by Zigang Geng (zigang@mail.ustc.edu.cn).
Used for: the DEKR-W32-NO-DC architecture, its decode behaviour, and
          its evaluation preprocessing recipe, ported into
          libreyolo/models/dekr/nn.py, libreyolo/postprocess/dekr.py
          and libreyolo/preprocess/dekr.py.
          Source paths used, at that pin:
            src/super_gradients/training/models/pose_estimation_models/
              dekr_hrnet.py
            src/super_gradients/training/utils/pose_estimation/
              dekr_decode_callbacks.py
            src/super_gradients/training/transforms/keypoints/
              keypoints_longest_max_size.py
            src/super_gradients/training/transforms/keypoints/
              keypoints_pad_if_needed.py
            src/super_gradients/recipes/arch_params/
              pose_dekr_w32_no_dc_arch_params.yaml
          Published pretrained DEKR weights are served from the
          source provider's CDN. No per-artifact redistribution grant
          was found for them, so LibreYOLO links to that CDN and does
          not bundle or mirror the checkpoint.

--------------------------------------------------------------------
HRNet / DEKR and HigherHRNet (architecture lineage)
--------------------------------------------------------------------
Source: https://github.com/HRNet/DEKR
        pinned commit 7a303139e92bdf3eab8d899415ccac37374285a4
        https://github.com/HRNet/HigherHRNet-Human-Pose-Estimation
        pinned commit aa23881492ff511185acf756a2e14725cc4ab4d7
License: MIT License (both repositories)
Copyright (c) Microsoft Corporation
Used for: architecture cross-check only for the LibreYOLO `dekr`
          family. No code was copied from either repository; the
          original DEKR implementation uses deformable convolution
          and is not state-compatible with the no-DC model LibreYOLO
          ships.

--------------------------------------------------------------------
SuperGradients / PP-YOLOE
--------------------------------------------------------------------
Source: https://github.com/Deci-AI/super-gradients
        pinned commit 63de22c404d5740f34f7706c302b37fce3c8fe5d
License: Apache License 2.0 (repository LICENSE.md at that commit)
Copyright (c) 2021-2024 Deci AI
Used for: the PP-YOLOE architecture, decoder, postprocessing reference
          and training recipe, ported into libreyolo/models/ppyoloe/
          and libreyolo/postprocess/ppyoloe.py.
          Source paths used, at that pin:
            src/super_gradients/training/models/detection_models/
              csp_resnet.py
              pp_yolo_e/pp_yolo_e.py
              pp_yolo_e/pp_yolo_head.py
              pp_yolo_e/pan.py
              pp_yolo_e/post_prediction_callback.py
            src/super_gradients/modules/
              conv_bn_act_block.py
              repvgg_block.py
              se_blocks.py
            src/super_gradients/training/utils/bbox_utils.py
            src/super_gradients/recipes/arch_params/
              ppyoloe_arch_params.yaml
              ppyoloe_{s,m,l,x}_arch_params.yaml
            src/super_gradients/recipes/training_hyperparams/
              coco2017_ppyoloe_train_params.yaml
          The PP-YOLOE architecture was introduced by PaddleDetection
          (https://github.com/PaddlePaddle/PaddleDetection, Apache-2.0,
          pinned commit b25522a0f4bde8c80603f3ba5e3472059972e3b5) and
          the paper arXiv:2203.16250. PaddleDetection was consulted for
          semantic cross-checking only; no PaddleDetection code was
          copied.
          The pretrained COCO PP-YOLOE checkpoints are published on the
          source provider's CDN. This audit found no per-artifact
          license grant authorizing LibreYOLO to rehost them, so
          LibreYOLO does not mirror them: they are downloaded from that
          CDN on demand and SHA256-verified before loading. COCO
          training data is not bundled.

--------------------------------------------------------------------
SuperGradients / YOLO-NAS-R (rotated / OBB)
--------------------------------------------------------------------
Source: https://github.com/Deci-AI/super-gradients
        pull request 2014, pinned commit
        69141b55c1161d939939a270523a7eca5a645f72
License: Apache License 2.0 (repository LICENSE.md at that commit)
Copyright (c) 2021-2024 Deci AI
Used for: the YOLO-NAS-R rotated head, its decode, and the evaluation
          preprocessing recipe, ported into the existing YOLO-NAS
          family (libreyolo/models/yolonas/nn.py -- YoloNASRDFLHead,
          YoloNASRNDFLHeads, LibreYOLONASOBBModel;
          libreyolo/preprocess/yolonas.py; libreyolo/postprocess/yolonas.py).
          Source paths used, at that pin:
            src/super_gradients/training/models/detection_models/
              yolo_nas_r/yolo_nas_r_dfl_head.py
              yolo_nas_r/yolo_nas_r_ndfl_heads.py
              yolo_nas_r/yolo_nas_r_variants.py
              yolo_nas_r/yolo_nas_r_post_prediction_callback.py
            src/super_gradients/recipes/arch_params/
              yolo_nas_r_{s,m,l}_arch_params.yaml
          The pretrained DOTA2 YOLO-NAS-R weights are covered by a
          SEPARATE, non-redistributable license (LICENSE.YOLONAS-R.md
          at the same pin: no resale, sublicensing or distribution, and
          no commercial use). LibreYOLO does not mirror them; they are
          downloaded from Deci's CDN on demand and SHA256-verified
          before loading. See weights/LICENSE_NOTICE.txt.
          DOTA2 training data is not bundled and must be obtained from
          its own source under its own terms.

--------------------------------------------------------------------
EdgeCrafter
--------------------------------------------------------------------
Source: https://github.com/Intellindust-AI-Lab/EdgeCrafter
        (formerly EC-codehub/EdgeCrafter)
License: Apache License 2.0 for the earlier source used by the LibreEC port,
         not a statement about current upstream releases.
Used for: EC model family detection, segmentation, and pose architecture
          references (libreyolo/models/ec/).
License history: Apache-2.0 is preserved at upstream commit
          b9f31f8ce23ac112fac2b9b01bc6a1f25a016d92 (LICENSE).
          e0656d44b80ecb97ba96bb8a804a741169431e0f replaced it with the
          non-commercial EdgeCrafter License on August 20, 2026;
          124cb0922ebe4af55c025af448312c2374210815 only renamed LICENSE
          to LICENSE.md on August 24, 2026.
          These are license-history references, not a recovered source pin
          for the original port. See docs/provenance/edgecrafter.md.
Weights: Original COCO defaults retain their Apache-2.0 notices. Optional
         -obj2coco mirrors carry the non-commercial EdgeCrafter License.
         See weights/LICENSE_NOTICE.txt for the separate weight terms.

--------------------------------------------------------------------
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. The CUDA
          graph gradient-ownership step (libreyolo/training/cuda_graph.py)
          and the torch.compile compatibility changes in the RF-DETR
          DINOv2 backbone and transformer follow RF-DETR v1.11.0
          (commit d96d6ff9303ec5af7bb8f73e83f0f04ecb6ef728; PRs #1411, #1455,
          #1468).

--------------------------------------------------------------------
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.

--------------------------------------------------------------------
LingBot-Vision (Robbyant / Ant Group)
--------------------------------------------------------------------
Source: https://github.com/robbyant/lingbot-vision
Commit: 151e46321bae4399f8568829f190c7bdec216b49
License: Apache License 2.0
Copyright 2026 LingBot-Vision Contributors.
Used for: the LibreLingBotVision semantic family
          (libreyolo/models/lingbotvision/). The ViT backbone in nn.py is a
          native port of the upstream DINOv3-style Vision Transformer (axial
          RoPE over the patch grid, [CLS] plus 4 register tokens, LayerScale,
          masked-K-bias fused QKV); module and parameter names mirror the
          reference implementation so the Apache-2.0 backbone checkpoints
          load unchanged. Only the inference forward path is ported; the
          upstream self-supervised pretraining machinery is not. The dense
          head and training recipe are LibreYOLO's own (the upstream report's
          linear-probing protocol). Hosted LibreLingBotVision{s,b,l}-sem
          checkpoints combine the unchanged Apache-2.0 Robbyant backbones
          with an ADE20K head trained by LibreYOLO. Reference: "Vision
          Pretraining for Dense Spatial Perception", Fu et al., 2026,
          arXiv:2607.05247. See libreyolo/models/lingbotvision/NOTICE.

--------------------------------------------------------------------
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.

--------------------------------------------------------------------
QuickSRNet (Qualcomm AI Research)
--------------------------------------------------------------------
Architecture and checkpoint source: https://github.com/quic/aimet-model-zoo
Commit: 1bd2bf5b17cdda9251437c444009b29e1a25054b
Maintained integration source: https://github.com/qualcomm/ai-hub-models
Commit: 16dbeb5e2805d4ada7218026de72e36878717d46
License: BSD-3-Clause
Copyright (c) 2022 Qualcomm Innovation Center, Inc.
Copyright (c) 2025 Qualcomm Technologies, Inc. and/or its subsidiaries.
Citation: Choi, J. et al. "QuickSRNet: Plain Single-Image Super-Resolution
          Architecture for Faster Inference on Mobile Platforms." CVPRW 2023.
Used for: the QuickSRNet Medium 2x generator in
          libreyolo/models/quicksrnet/nn.py. Module names preserve exact
          compatibility with the official state dict. LibreYOLO adds the
          restore-task wrapper, native-resolution RGB preprocessing,
          scale-aware postprocessing, PSNR/SSIM validation, dynamic spatial
          ONNX export, checkpoint conversion, and tests.

NOTE - code vs. weights/data: the architecture code and the official Medium
2x checkpoint are distributed by the BSD-3-Clause AIMET Model Zoo project.
The checkpoint was trained on DIV2K. LibreYOLO's source distribution does not
bundle the checkpoint. The separate Hugging Face mirror copies the 14 learned
tensors unchanged, discards optimizer/history objects, adds LibreYOLO metadata,
and ships the verbatim upstream BSD-3-Clause license. DIV2K is training-data
provenance, not part of LibreYOLO or the mirrored checkpoint repository.

--------------------------------------------------------------------
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 fp16/fp8
          variants). No code was copied from the nobg library. Upstream
          training data is not disclosed; LibreYOLO redistributes only
          the released weights.

--------------------------------------------------------------------
BEN2 (Prama LLC)
--------------------------------------------------------------------
Source: https://github.com/PramaLLC/BEN2
Commit: 2c99a5da477b5523585bfa5c893888a6e818a8f6
License: MIT
Copyright (c) 2025 Prama LLC
Used for: the BEN2 Base background-removal model family
          (libreyolo/models/ben2/, matte task). The port covers the
          public Base mask network, keeps the released parameter names,
          replaces einops-only reshapes with native PyTorch, removes
          global random-state mutations, and returns logits for the shared
          matte-task postprocessing contract. Exact fp32 parity is recorded
          by weights/parity_ben2.py for batch 1 and batch 2 at 1024.

BEN2's multi-field cross-attention and refinement modules follow MVANet:
Source: https://github.com/qianyu-dlut/MVANet
Commit: edecec5db30c1c40d5204c630ec3779a428795ee
License: MIT
Copyright (c) 2024 qianyu-dlut

The optional upstream foreground-colour refinement helper is excluded because
its separately attributed source has no declared license. Upstream video and
media helpers are also outside the port. Released BEN2 Base weights are
declared MIT by the upstream Hugging Face model card and are documented in
weights/LICENSE_NOTICE.txt. Upstream states that BEN2 was trained on DIS5K and
Prama LLC's proprietary 22K segmentation dataset; LibreYOLO redistributes no
training data.

--------------------------------------------------------------------
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 by these V1 ports; ConvNeXt V2 is listed separately below.

          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/tree/4998cec385b1aaa07abdefba71bfba2f83c7db32
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 v1.0.1 release by libreyolo/models/sam3dbody/mhr_body.py and cached
locally; it is NOT mirrored by LibreYOLO. The release archive SHA-256 is
e4f4f205cd87c0fa106577ba1de4fc763e4eb197c924461d2ef7e6944e9d6b94 and the
extracted mhr_model.pt SHA-256 is
352e271a6c42729c68554ceaea0c955e866970160c31e35506d782dc0f7377bc. 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/tree/b5c765a0d89d789985e186d396315e7590887b94
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 automatic access gate recording license acceptance plus
self-attested identity and jurisdiction. The gate is not manual approval or
screening by Meta. Mirroring does not change the license: those weights remain
SAM-licensed, not MIT. Runtime transport is pinned respectively to Hub commits
46e286e25347518d861ab0f21e1b2b5b630dc21f and
a745fa6fcd5d71e16c4da921a28a6bb6f1ff9e3e and verifies the complete local
runtime file set immediately before and after the SAM-licensed path-only loader
runs. That handoff requires a trusted, quiescent same-user local filesystem;
the wrapper does not claim to prevent transient substitution by a process that
can mutate the same paths between the two checks.

--------------------------------------------------------------------
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.

--------------------------------------------------------------------
ShowUI grounding system instruction (Show Lab)
--------------------------------------------------------------------
Source: https://huggingface.co/showlab/ShowUI-2B
Commit: cabec4fcc48d15ffd3efe0b33ea9bc7d41509d60
License: MIT (the pinned model card declares ``license: mit``)
Copyright (c) 2024 Show Lab
Used for: the verbatim grounding system instruction in
          libreyolo/models/ground/showui.py. The adapter control flow is
          original LibreYOLO code; no upstream model implementation is
          vendored or ported. The license text below is preserved verbatim
          from showlab/ShowUI commit
          59e059c4df62db0857cc7a7ef0b15d067a30c274.

MIT License

Copyright (c) 2024 Show 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.

--------------------------------------------------------------------
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.

--------------------------------------------------------------------
WildDet3D (optional external runtime)
--------------------------------------------------------------------
LibreWildDet3D is an independently written MIT API adapter.

Optional runtime: https://github.com/allenai/WildDet3D
Public-interface reference revision: 1b8aa52b6ff3f00d0ebfa07175efc0c0c440964a
Reference: docs/INFERENCE.md (public API and tensor-format documentation).
Runtime license: SAM License (custom, field-of-use restrictions).
No upstream model, preprocessing, postprocessing, or rendering implementation
is copied, adapted, or vendored. The user installs the optional runtime
separately. The adapter calls build_model, preprocess, and the predictor API.

Runtime submodules at that revision:
- sam3_for_wilddet3d: 159490bb13108749dd5ceff68f42cd1792e60641, SAM License.
- LingBot-Depth: ff29de5bf433d1b1519cd851da00bc652668964a, Apache-2.0.
Their files are not distributed by LibreYOLO.
The Stage 3 checkpoint is mirrored byte-identically at
https://huggingface.co/LibreYOLO/LibreWildDet3D revision
e2baa17eb2279225094a2d7e7d7620ff06dc4676. It retains the SAM License and its
original filename and serialization. LibreYOLO verifies SHA-256
f8b6a9e548f733ba62625a0d2adc4b0f4fdb6007ee11d9927f9c1027010fee57 after
automatic download. Explicit local upstream checkpoints remain supported.

--------------------------------------------------------------------
3D-MOOD (optional external runtime)
--------------------------------------------------------------------
Source: https://github.com/cvg/3D-MOOD
Revision: 41bb2904932d91507338e75ae4c802d67616ca1b
License: Apache-2.0 (text in licenses/Apache-2.0.txt)
Copyright 2025 Computer Vision and Geometry Lab, ETH Zurich.
Used for: the model-construction, preprocessing, checkpoint-loading, and output
          contracts implemented by the adapter and private worker in
          libreyolo/models/mood3d/. The upstream runtime is separately installed.

The upstream Swin implementation attributes OpenMMLab/mmdetection (Apache-2.0),
its Grounding DINO modules attribute IDEA-Research/GroundingDINO (Apache-2.0),
its deformable-attention implementation attributes fundamentalvision/
Deformable-DETR (Apache-2.0), and its geometry utilities attribute PyTorch3D
(BSD-3-Clause). LibreYOLO's worker uses its existing Apache-2.0 portable
deformable-attention implementation. It does not use SysCV/vis4d_cuda_ops,
whose repository has no declared license.

--------------------------------------------------------------------
vis4d cuboid geometry convention
--------------------------------------------------------------------
Source: https://github.com/SysCV/vis4d
Revision: 9cd052ed72a208773abc1f6f0cd9109d8ec97531 (v1.0.0)
License: Apache-2.0 (text in licenses/Apache-2.0.txt)
Copyright 2021 Visual Intelligence and Systems (VIS) Group at ETH Zurich.
Used for: OpenCV cuboid dimension-to-axis mapping and corner ordering in
libreyolo/utils/results.py::Boxes3D.corners and the associated wireframe edges.
Reference: vis4d/op/box/box3d.py::boxes3d_to_corners, AxisMode.OPENCV.
The implementation remains numpy-based; vis4d is not a core dependency.


FCOS3D inference implementation

Derived from OpenMMLab, Apache License 2.0:
- https://github.com/open-mmlab/mmdetection3d
  Revision: fe25f7a51d36e3702f961e198894580d83c4387b
  Source: FCOSMono3DHead, AnchorFreeMono3DHead, FCOS3DBBoxCoder, FCOS3D configs.
- https://github.com/open-mmlab/mmdetection
  Revision: cfd5d3a985b0249de009b67d04f37263e11cdf3d
  Source: ResNet, ResLayer and FPN.
Copyright (c) OpenMMLab. All rights reserved.

LibreYOLO changes: standalone PyTorch network, no framework registry or
training dependencies, torchvision deform_conv2d, camera Results mapping.
The torchvision operator is a separately installed BSD-3-Clause dependency.
No third-party C++/CUDA operator source is bundled.

This notice covers code, not the official nuScenes checkpoint. No weights are
bundled in this port. The official checkpoint is mirrored byte-for-byte at
LibreYOLO/LibreFCOS3D revision af69c4abfb265f63695ecbaa68b4167f0c150605.
OpenMMLab declares no per-object checkpoint license; LibreYOLO's maintainer
approved that mirror by treating the project's Apache-2.0 statement as the
redistribution basis, which is LibreYOLO's disclosed interpretation and not a
clarification from the OpenMMLab authors. The weights are trained on nuScenes
under non-commercial terms (https://www.nuscenes.org/terms-of-use) and are not
covered by LibreYOLO's MIT license.

Copyright 2018-2019 Open-MMLab. All rights reserved.

                                 Apache License
                           Version 2.0, January 2004
                        http://www.apache.org/licenses/

   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION

   1. Definitions.

      "License" shall mean the terms and conditions for use, reproduction,
      and distribution as defined by Sections 1 through 9 of this document.

      "Licensor" shall mean the copyright owner or entity authorized by
      the copyright owner that is granting the License.

      "Legal Entity" shall mean the union of the acting entity and all
      other entities that control, are controlled by, or are under common
      control with that entity. For the purposes of this definition,
      "control" means (i) the power, direct or indirect, to cause the
      direction or management of such entity, whether by contract or
      otherwise, or (ii) ownership of fifty percent (50%) or more of the
      outstanding shares, or (iii) beneficial ownership of such entity.

      "You" (or "Your") shall mean an individual or Legal Entity
      exercising permissions granted by this License.

      "Source" form shall mean the preferred form for making modifications,
      including but not limited to software source code, documentation
      source, and configuration files.

      "Object" form shall mean any form resulting from mechanical
      transformation or translation of a Source form, including but
      not limited to compiled object code, generated documentation,
      and conversions to other media types.

      "Work" shall mean the work of authorship, whether in Source or
      Object form, made available under the License, as indicated by a
      copyright notice that is included in or attached to the work
      (an example is provided in the Appendix below).

      "Derivative Works" shall mean any work, whether in Source or Object
      form, that is based on (or derived from) the Work and for which the
      editorial revisions, annotations, elaborations, or other modifications
      represent, as a whole, an original work of authorship. For the purposes
      of this License, Derivative Works shall not include works that remain
      separable from, or merely link (or bind by name) to the interfaces of,
      the Work and Derivative Works thereof.

      "Contribution" shall mean any work of authorship, including
      the original version of the Work and any modifications or additions
      to that Work or Derivative Works thereof, that is intentionally
      submitted to Licensor for inclusion in the Work by the copyright owner
      or by an individual or Legal Entity authorized to submit on behalf of
      the copyright owner. For the purposes of this definition, "submitted"
      means any form of electronic, verbal, or written communication sent
      to the Licensor or its representatives, including but not limited to
      communication on electronic mailing lists, source code control systems,
      and issue tracking systems that are managed by, or on behalf of, the
      Licensor for the purpose of discussing and improving the Work, but
      excluding communication that is conspicuously marked or otherwise
      designated in writing by the copyright owner as "Not a Contribution."

      "Contributor" shall mean Licensor and any individual or Legal Entity
      on behalf of whom a Contribution has been received by Licensor and
      subsequently incorporated within the Work.

   2. Grant of Copyright License. Subject to the terms and conditions of
      this License, each Contributor hereby grants to You a perpetual,
      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
      copyright license to reproduce, prepare Derivative Works of,
      publicly display, publicly perform, sublicense, and distribute the
      Work and such Derivative Works in Source or Object form.

   3. Grant of Patent License. Subject to the terms and conditions of
      this License, each Contributor hereby grants to You a perpetual,
      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
      (except as stated in this section) patent license to make, have made,
      use, offer to sell, sell, import, and otherwise transfer the Work,
      where such license applies only to those patent claims licensable
      by such Contributor that are necessarily infringed by their
      Contribution(s) alone or by combination of their Contribution(s)
      with the Work to which such Contribution(s) was submitted. If You
      institute patent litigation against any entity (including a
      cross-claim or counterclaim in a lawsuit) alleging that the Work
      or a Contribution incorporated within the Work constitutes direct
      or contributory patent infringement, then any patent licenses
      granted to You under this License for that Work shall terminate
      as of the date such litigation is filed.

   4. Redistribution. You may reproduce and distribute copies of the
      Work or Derivative Works thereof in any medium, with or without
      modifications, and in Source or Object form, provided that You
      meet the following conditions:

      (a) You must give any other recipients of the Work or
          Derivative Works a copy of this License; and

      (b) You must cause any modified files to carry prominent notices
          stating that You changed the files; and

      (c) You must retain, in the Source form of any Derivative Works
          that You distribute, all copyright, patent, trademark, and
          attribution notices from the Source form of the Work,
          excluding those notices that do not pertain to any part of
          the Derivative Works; and

      (d) If the Work includes a "NOTICE" text file as part of its
          distribution, then any Derivative Works that You distribute must
          include a readable copy of the attribution notices contained
          within such NOTICE file, excluding those notices that do not
          pertain to any part of the Derivative Works, in at least one
          of the following places: within a NOTICE text file distributed
          as part of the Derivative Works; within the Source form or
          documentation, if provided along with the Derivative Works; or,
          within a display generated by the Derivative Works, if and
          wherever such third-party notices normally appear. The contents
          of the NOTICE file are for informational purposes only and
          do not modify the License. You may add Your own attribution
          notices within Derivative Works that You distribute, alongside
          or as an addendum to the NOTICE text from the Work, provided
          that such additional attribution notices cannot be construed
          as modifying the License.

      You may add Your own copyright statement to Your modifications and
      may provide additional or different license terms and conditions
      for use, reproduction, or distribution of Your modifications, or
      for any such Derivative Works as a whole, provided Your use,
      reproduction, and distribution of the Work otherwise complies with
      the conditions stated in this License.

   5. Submission of Contributions. Unless You explicitly state otherwise,
      any Contribution intentionally submitted for inclusion in the Work
      by You to the Licensor shall be under the terms and conditions of
      this License, without any additional terms or conditions.
      Notwithstanding the above, nothing herein shall supersede or modify
      the terms of any separate license agreement you may have executed
      with Licensor regarding such Contributions.

   6. Trademarks. This License does not grant permission to use the trade
      names, trademarks, service marks, or product names of the Licensor,
      except as required for reasonable and customary use in describing the
      origin of the Work and reproducing the content of the NOTICE file.

   7. Disclaimer of Warranty. Unless required by applicable law or
      agreed to in writing, Licensor provides the Work (and each
      Contributor provides its Contributions) on an "AS IS" BASIS,
      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
      implied, including, without limitation, any warranties or conditions
      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
      PARTICULAR PURPOSE. You are solely responsible for determining the
      appropriateness of using or redistributing the Work and assume any
      risks associated with Your exercise of permissions under this License.

   8. Limitation of Liability. In no event and under no legal theory,
      whether in tort (including negligence), contract, or otherwise,
      unless required by applicable law (such as deliberate and grossly
      negligent acts) or agreed to in writing, shall any Contributor be
      liable to You for damages, including any direct, indirect, special,
      incidental, or consequential damages of any character arising as a
      result of this License or out of the use or inability to use the
      Work (including but not limited to damages for loss of goodwill,
      work stoppage, computer failure or malfunction, or any and all
      other commercial damages or losses), even if such Contributor
      has been advised of the possibility of such damages.

   9. Accepting Warranty or Additional Liability. While redistributing
      the Work or Derivative Works thereof, You may choose to offer,
      and charge a fee for, acceptance of support, warranty, indemnity,
      or other liability obligations and/or rights consistent with this
      License. However, in accepting such obligations, You may act only
      on Your own behalf and on Your sole responsibility, not on behalf
      of any other Contributor, and only if You agree to indemnify,
      defend, and hold each Contributor harmless for any liability
      incurred by, or claims asserted against, such Contributor by reason
      of your accepting any such warranty or additional liability.

   END OF TERMS AND CONDITIONS

   APPENDIX: How to apply the Apache License to your work.

      To apply the Apache License to your work, attach the following
      boilerplate notice, with the fields enclosed by brackets "[]"
      replaced with your own identifying information. (Don't include
      the brackets!)  The text should be enclosed in the appropriate
      comment syntax for the file format. We also recommend that a
      file or class name and description of purpose be included on the
      same "printed page" as the copyright notice for easier
      identification within third-party archives.

   Copyright 2018-2019 Open-MMLab.

   Licensed under the Apache License, Version 2.0 (the "License");
   you may not use this file except in compliance with the License.
   You may obtain a copy of the License at

       http://www.apache.org/licenses/LICENSE-2.0

   Unless required by applicable law or agreed to in writing, software
   distributed under the License is distributed on an "AS IS" BASIS,
   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
   See the License for the specific language governing permissions and
   limitations under the License.


DetAny3D runtime adapter

The worker input recipe and decoding interface are adapted from:
https://github.com/OpenDriveLab/DetAny3D
Revision: 10e484be0837d80aa33cff6ed95a9be834a3f794
Apache License 2.0. Source: deploy.py, wrap_model.py and train_utils.py.
LibreYOLO changes: typed-array worker boundary, CPU compatibility, original-image
coordinate mapping, canonical cuboid axes, input validation and CLI integration.

CPU attention compatibility implements the documented xFormers 0.0.16 API:
https://github.com/facebookresearch/xformers
Revision: 6f3c20f223c0f051e4b2dc744cc7591591ee04c4
BSD 3-Clause. Copyright (c) Facebook, Inc. and its affiliates.
The xFormers Python package is separately installed. No xFormers kernels are
bundled. Deformable attention reuses LibreYOLO's existing Apache-2.0 implementation.

The model implementation and checkpoint are not bundled. The separately
installed upstream runtime includes UniDepth under CC BY-NC 4.0:
https://github.com/lpiccinelli-eth/UniDepth
Revision checked: 8d8cfe4c7ee15297099983607febf0d4f32eb3d6
No UniDepth implementation was used to write or copied into this adapter.
This does not grant commercial-use rights to the external runtime.
The official full checkpoint is mirrored byte-for-byte at
LibreYOLO/LibreDetAny3D revision eeb89d1b1d37ec5a361b1ec79b26ea0620d9f4c3.
OpenDriveLab declares no per-object checkpoint license; LibreYOLO's maintainer
approved that mirror by treating the project's Apache-2.0 statement as the
redistribution basis for the parameters DetAny3D itself trained. This is
LibreYOLO's disclosed interpretation and not a clarification from the DetAny3D
authors. Upstream documents the checkpoint's depth branch as initialized from
UniDepth v2, so LibreYOLO applies CC BY-NC 4.0 to the mirrored file as the
strictest known term. The weights are non-commercial and are not covered by
LibreYOLO's MIT license.

                                 Apache License
                           Version 2.0, January 2004
                        http://www.apache.org/licenses/

   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION

   1. Definitions.

      "License" shall mean the terms and conditions for use, reproduction,
      and distribution as defined by Sections 1 through 9 of this document.

      "Licensor" shall mean the copyright owner or entity authorized by
      the copyright owner that is granting the License.

      "Legal Entity" shall mean the union of the acting entity and all
      other entities that control, are controlled by, or are under common
      control with that entity. For the purposes of this definition,
      "control" means (i) the power, direct or indirect, to cause the
      direction or management of such entity, whether by contract or
      otherwise, or (ii) ownership of fifty percent (50%) or more of the
      outstanding shares, or (iii) beneficial ownership of such entity.

      "You" (or "Your") shall mean an individual or Legal Entity
      exercising permissions granted by this License.

      "Source" form shall mean the preferred form for making modifications,
      including but not limited to software source code, documentation
      source, and configuration files.

      "Object" form shall mean any form resulting from mechanical
      transformation or translation of a Source form, including but
      not limited to compiled object code, generated documentation,
      and conversions to other media types.

      "Work" shall mean the work of authorship, whether in Source or
      Object form, made available under the License, as indicated by a
      copyright notice that is included in or attached to the work
      (an example is provided in the Appendix below).

      "Derivative Works" shall mean any work, whether in Source or Object
      form, that is based on (or derived from) the Work and for which the
      editorial revisions, annotations, elaborations, or other modifications
      represent, as a whole, an original work of authorship. For the purposes
      of this License, Derivative Works shall not include works that remain
      separable from, or merely link (or bind by name) to the interfaces of,
      the Work and Derivative Works thereof.

      "Contribution" shall mean any work of authorship, including
      the original version of the Work and any modifications or additions
      to that Work or Derivative Works thereof, that is intentionally
      submitted to Licensor for inclusion in the Work by the copyright owner
      or by an individual or Legal Entity authorized to submit on behalf of
      the copyright owner. For the purposes of this definition, "submitted"
      means any form of electronic, verbal, or written communication sent
      to the Licensor or its representatives, including but not limited to
      communication on electronic mailing lists, source code control systems,
      and issue tracking systems that are managed by, or on behalf of, the
      Licensor for the purpose of discussing and improving the Work, but
      excluding communication that is conspicuously marked or otherwise
      designated in writing by the copyright owner as "Not a Contribution."

      "Contributor" shall mean Licensor and any individual or Legal Entity
      on behalf of whom a Contribution has been received by Licensor and
      subsequently incorporated within the Work.

   2. Grant of Copyright License. Subject to the terms and conditions of
      this License, each Contributor hereby grants to You a perpetual,
      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
      copyright license to reproduce, prepare Derivative Works of,
      publicly display, publicly perform, sublicense, and distribute the
      Work and such Derivative Works in Source or Object form.

   3. Grant of Patent License. Subject to the terms and conditions of
      this License, each Contributor hereby grants to You a perpetual,
      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
      (except as stated in this section) patent license to make, have made,
      use, offer to sell, sell, import, and otherwise transfer the Work,
      where such license applies only to those patent claims licensable
      by such Contributor that are necessarily infringed by their
      Contribution(s) alone or by combination of their Contribution(s)
      with the Work to which such Contribution(s) was submitted. If You
      institute patent litigation against any entity (including a
      cross-claim or counterclaim in a lawsuit) alleging that the Work
      or a Contribution incorporated within the Work constitutes direct
      or contributory patent infringement, then any patent licenses
      granted to You under this License for that Work shall terminate
      as of the date such litigation is filed.

   4. Redistribution. You may reproduce and distribute copies of the
      Work or Derivative Works thereof in any medium, with or without
      modifications, and in Source or Object form, provided that You
      meet the following conditions:

      (a) You must give any other recipients of the Work or
          Derivative Works a copy of this License; and

      (b) You must cause any modified files to carry prominent notices
          stating that You changed the files; and

      (c) You must retain, in the Source form of any Derivative Works
          that You distribute, all copyright, patent, trademark, and
          attribution notices from the Source form of the Work,
          excluding those notices that do not pertain to any part of
          the Derivative Works; and

      (d) If the Work includes a "NOTICE" text file as part of its
          distribution, then any Derivative Works that You distribute must
          include a readable copy of the attribution notices contained
          within such NOTICE file, excluding those notices that do not
          pertain to any part of the Derivative Works, in at least one
          of the following places: within a NOTICE text file distributed
          as part of the Derivative Works; within the Source form or
          documentation, if provided along with the Derivative Works; or,
          within a display generated by the Derivative Works, if and
          wherever such third-party notices normally appear. The contents
          of the NOTICE file are for informational purposes only and
          do not modify the License. You may add Your own attribution
          notices within Derivative Works that You distribute, alongside
          or as an addendum to the NOTICE text from the Work, provided
          that such additional attribution notices cannot be construed
          as modifying the License.

      You may add Your own copyright statement to Your modifications and
      may provide additional or different license terms and conditions
      for use, reproduction, or distribution of Your modifications, or
      for any such Derivative Works as a whole, provided Your use,
      reproduction, and distribution of the Work otherwise complies with
      the conditions stated in this License.

   5. Submission of Contributions. Unless You explicitly state otherwise,
      any Contribution intentionally submitted for inclusion in the Work
      by You to the Licensor shall be under the terms and conditions of
      this License, without any additional terms or conditions.
      Notwithstanding the above, nothing herein shall supersede or modify
      the terms of any separate license agreement you may have executed
      with Licensor regarding such Contributions.

   6. Trademarks. This License does not grant permission to use the trade
      names, trademarks, service marks, or product names of the Licensor,
      except as required for reasonable and customary use in describing the
      origin of the Work and reproducing the content of the NOTICE file.

   7. Disclaimer of Warranty. Unless required by applicable law or
      agreed to in writing, Licensor provides the Work (and each
      Contributor provides its Contributions) on an "AS IS" BASIS,
      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
      implied, including, without limitation, any warranties or conditions
      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
      PARTICULAR PURPOSE. You are solely responsible for determining the
      appropriateness of using or redistributing the Work and assume any
      risks associated with Your exercise of permissions under this License.

   8. Limitation of Liability. In no event and under no legal theory,
      whether in tort (including negligence), contract, or otherwise,
      unless required by applicable law (such as deliberate and grossly
      negligent acts) or agreed to in writing, shall any Contributor be
      liable to You for damages, including any direct, indirect, special,
      incidental, or consequential damages of any character arising as a
      result of this License or out of the use or inability to use the
      Work (including but not limited to damages for loss of goodwill,
      work stoppage, computer failure or malfunction, or any and all
      other commercial damages or losses), even if such Contributor
      has been advised of the possibility of such damages.

   9. Accepting Warranty or Additional Liability. While redistributing
      the Work or Derivative Works thereof, You may choose to offer,
      and charge a fee for, acceptance of support, warranty, indemnity,
      or other liability obligations and/or rights consistent with this
      License. However, in accepting such obligations, You may act only
      on Your own behalf and on Your sole responsibility, not on behalf
      of any other Contributor, and only if You agree to indemnify,
      defend, and hold each Contributor harmless for any liability
      incurred by, or claims asserted against, such Contributor by reason
      of your accepting any such warranty or additional liability.

   END OF TERMS AND CONDITIONS

   APPENDIX: How to apply the Apache License to your work.

      To apply the Apache License to your work, attach the following
      boilerplate notice, with the fields enclosed by brackets "[]"
      replaced with your own identifying information. (Don't include
      the brackets!)  The text should be enclosed in the appropriate
      comment syntax for the file format. We also recommend that a
      file or class name and description of purpose be included on the
      same "printed page" as the copyright notice for easier
      identification within third-party archives.

   Copyright [yyyy] [name of copyright owner]

   Licensed under the Apache License, Version 2.0 (the "License");
   you may not use this file except in compliance with the License.
   You may obtain a copy of the License at

       http://www.apache.org/licenses/LICENSE-2.0

   Unless required by applicable law or agreed to in writing, software
   distributed under the License is distributed on an "AS IS" BASIS,
   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
   See the License for the specific language governing permissions and
   limitations under the License.

--------------------------------------------------------------------
Marigold V2
Copyright 2026 Huawei Technologies Co., Ltd.

Inference graph and preprocessing adapted from:
https://github.com/huawei-bayerlab/marigold-v2
Revision: cc6a7031abcd59fd9e1ceff7fdd0d9687d389bc5
License: Apache-2.0 (LICENSE alongside this notice).
Sources: experiments/20260316_qwen_depth/component_loader.py and network_graph.py,
dataset/dataloading/util.py and transform.py, validation/folder_steps.py and
validate_steps.py. Paths above are relative to upstream marigoldv2/.
LibreYOLO changes: registry-free graph, strict adapter loading, tensor-only
prompt deserialization, checkpoint metadata, isolated inference RNG state,
original-canvas results and explicit depth/normal/albedo conventions.

The original upstream NOTICE credits Marigold (https://github.com/prs-eth/Marigold),
Copyright 2023-2025 Marigold Team, ETH Zurich, Apache-2.0, for reused evaluation,
training and utility code. Those evaluation/training implementations are not
bundled by this family.

Architecture implementations load from separately installed dependencies:
huggingface/diffusers v0.38.0 (Apache-2.0), huggingface/peft v0.18.1
(Apache-2.0), bitsandbytes-foundation/bitsandbytes 0.49.2 (MIT).
No DINOv3 teacher implementation is bundled or used in inference.

Pretrained adapter and prompt tensors are sourced separately from:
https://huggingface.co/huawei-bayerlab/marigold-v2-0
Revision: 6fd6d1ca246c9d2d99a4d8ac375a4eccc87178ad
License: Apache-2.0, as declared in that model repository's LICENSE.
Copyright 2026 Huawei Technologies Co., Ltd.
The converted checkpoints retain inference and prompt tensors unchanged and
omit two named training-only iREPA projection tensors.

The frozen base is downloaded separately from:
https://huggingface.co/Qwen/Qwen-Image-Edit-2509
Revision: d3968ef930e841f4c73640fb8afa3b306a78167e
License: Apache-2.0, as declared by the base model card.

--------------------------------------------------------------------
Molmo2 single-image pointing adapter
-----------------------------------

Upstream: Allen Institute for AI (Ai2), https://github.com/allenai/molmo2
Commit: f3cb1085fbb97c4a4d7fdcadd77cf871bf37a88a
License: Apache-2.0 (full text: licenses/Apache-2.0.txt in LibreYOLO)

The adapter and parser integrate the upstream Hugging Face usage API and the
point grammar documented in olmo/preprocessing/point_formatter.py and its tests.
LibreYOLO reuses its existing VLM inference and point-result implementations;
no Molmo2 architecture source is bundled in this Python package.
Remote model/processor code and weights originate from these Apache-2.0
snapshots, with unmodified learned parameters:

allenai/Molmo2-4B   @ 042abfa7a38879a376cec03d949eff0aefaa0600
allenai/Molmo2-8B   @ e28fa28597e5ec5e0cca2201dd8ab33d48bc4a1b
allenai/Molmo2-O-7B @ 784410650d12be9bc086118fdefa32d2c3bced86

4B/8B use Qwen3 language backbones; O-7B uses OLMo 3. The released snapshots
include the SigLIP vision backbone and their tokenizer assets. Upstream model
cards declare Apache-2.0 for all three snapshots.


GTR (Gated Token Recurrence)
Source: https://github.com/Intellindust-AI-Lab/GTR
Revision: 782e737efe2e6437ac537fbdcee089673d3376c1
License: MIT (LICENSE in this directory)
Copyright (c) 2026 Intellindust-AI-Lab / The GTR Authors.

Ported backbone, spatial SwiGLU, feature projector, decoder, matching, loss,
scheduler, Mosaic and batch MixUp, and the depth head and SILog loss in
libreyolo/models/gtr/.
GTRSeg instance-segmentation model (configs/seg) in libreyolo/models/gtr/.
The pose decoder reuses LibreYOLO's ECPose port (libreyolo/models/ec/) and
adapts GTR's pose decoder layer in libreyolo/models/gtr/pose.py.
Ported the oriented-box (OBB) decoder, rotated-box operations, oriented
denoising, criterion and matcher from engine/gtr/obb/, and the OBB flip and
rotation augmentation from engine/data/transforms/obb_transforms.py, as
libreyolo/models/gtr/obb_*.py.
Semantic segmentation head, loss and sliding-window evaluation
(engine/gtr/semseg, engine/solver/semseg_solver.py) in sem.py and
sem_trainer.py; Cityscapes weights from Phoenix8125/GTR semseg/.
Modified for native LibreYOLO construction, portable attention, local imports,
strict checkpoints, and validation loss output/normalization.

Retained source notices identify these Apache-2.0 components:
- Depth-Anything-V2 (DPT depth head, SILog loss; code only):
  https://github.com/DepthAnything/Depth-Anything-V2
  Copyright (c) 2024 Depth Anything V2 authors.
- DEIM (Mosaic, batch MixUp): https://github.com/Intellindust-AI-Lab/DEIM
  Copyright (c) 2024 The DEIM Authors.
- DETRPose (pose decoder): https://github.com/SebastianJanampa/DETRPose
  Copyright (c) the DETRPose authors.
- Conditional DETR (pose decoder): https://github.com/Atten4Vis/ConditionalDETR
  Copyright (c) 2021 Microsoft.
- D-FINE: https://github.com/Peterande/D-FINE
  Copyright (c) 2024 D-FINE Authors.
- RT-DETR: https://github.com/lyuwenyu/RT-DETR
  Copyright (c) 2023 lyuwenyu.
- DETR: https://github.com/facebookresearch/detr
  Copyright (c) Facebook, Inc. and its affiliates.
- RF-DETR segmentation helpers used by the upstream decoder/criterion:
  https://github.com/roboflow/rf-detr
  Copyright (c) 2025 Roboflow.
Their upstream-derived portions retain Apache-2.0; see LICENSE-APACHE.md.

Optional flash-linear-attention runtime acceleration is not vendored.
Source: https://github.com/fla-org/flash-linear-attention (MIT).
The portable recurrence is implemented directly in attention.py.

MIT License

Copyright (c) 2026 Intellindust-AI-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.
