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.

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

MIT License

Copyright (c) 2020 obss

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

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

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

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

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

MIT License

Copyright (c) 2021 OpenAI

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

--------------------------------------------------------------------
HuggingFace Transformers
--------------------------------------------------------------------
Source: https://github.com/huggingface/transformers
License: Apache License 2.0
Copyright 2022-2024 The HuggingFace Team. All Rights Reserved.
Used for: DINOv2-with-Registers reference implementation that
          libreyolo/models/rfdetr/dinov2.py adapts to add windowed
          self-attention. Also a runtime dependency loaded via
          AutoBackbone for the non-windowed DinoV2 path.

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

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

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

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

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

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

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

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

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

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

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

--------------------------------------------------------------------
LW-DETR (Atten4Vis / Baidu)
--------------------------------------------------------------------
Source: https://github.com/Atten4Vis/LW-DETR
License: Apache License 2.0
Copyright (c) 2024 Baidu. All Rights Reserved.
Used for: backbone, transformer, matcher, loss, postprocess, and
          tensor utilities consumed by RF-DETR
          (libreyolo/models/rfdetr/{backbone,transformer,matcher,
          loss,lwdetr,tensors,box_ops}.py).

--------------------------------------------------------------------
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).

--------------------------------------------------------------------
DETR (facebookresearch / Meta)
--------------------------------------------------------------------
Source: https://github.com/facebookresearch/detr
License: Apache License 2.0
Copyright (c) Facebook, Inc. and its affiliates.
Used for: 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).

--------------------------------------------------------------------
Deformable DETR (fundamentalvision / SenseTime)
--------------------------------------------------------------------
Source: https://github.com/fundamentalvision/Deformable-DETR
License: Apache License 2.0
Copyright (c) 2020 SenseTime. All Rights Reserved.
Used for: multi-scale deformable attention reused by RF-DETR
          (libreyolo/models/rfdetr/transformer.py: MSDeformAttn,
          ms_deform_attn_core_pytorch).

--------------------------------------------------------------------
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).

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

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

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

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

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

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

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

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

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

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

--------------------------------------------------------------------
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,
          efficientnetv2,resnet}/ — 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. Architecture lineage: ConvNeXt also
          derives from facebookresearch/ConvNeXt (MIT); EfficientNetV2 from
          google/automl (Apache-2.0); ResNet from He et al. 2015. Weights
          (timm *.in1k / *.fb_in1k / a1_in1k, Apache-2.0) are mirrored on
          the LibreYOLO Hugging Face org. ConvNeXt-V2 fcmae weights
          (CC-BY-NC) are NOT 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.
--------------------------------------------------------------------
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.
