NOTICE: WEIGHTS LICENSING

No pretrained weight files are distributed in this directory or anywhere
else in the source tree. The conversion scripts in this directory are
part of the LibreYOLO source repository and are covered by the root
MIT License.

Pretrained LibreYOLO weights are published separately on Hugging Face
under the LibreYOLO organization:

    https://huggingface.co/LibreYOLO

Each Hugging Face model repository ships its own LICENSE, NOTICE (where
applicable), and attribution reflecting the license of the upstream
project the weights were derived from. Per-family summary:

    YOLOX    (LibreYOLOX{n,t,s,m,l,x})                 Apache-2.0
             upstream: Megvii-BaseDetection/YOLOX

    YOLO9    (LibreYOLO9{t,s,m,c})                     MIT
             upstream: MultimediaTechLab/YOLO

    YOLOv1   (LibreYOLO1{t,b})                         public domain
    YOLOv2   (LibreYOLO2{t,b})                         public domain
    YOLOv3   (LibreYOLO3{t,b,spp})                     public domain
    YOLOv4   (LibreYOLO4{t,b})                         public domain
             upstream: pjreddie/darknet (v1/v2/v3), AlexeyAB/darknet (v4)
             Darknet is public domain ("YOLO LICENSE"); both the .cfg
             architectures and the pretrained .weights carry no license
             obligations. Convert with weights/convert_darknet_weights.py.
             YOLOv1 is Pascal VOC (20 classes), fixed 448. Its pretrained
             yolov1.weights was pulled from pjreddie.com/media/files/ (now
             archived only on the Wayback Machine); the tiny-yolov1 .weights
             was overwritten upstream by the YOLOv2 release and is not
             recoverable, so LibreYOLO1t ships as code only (bring your own
             tiny-yolov1.weights to convert).

    Dome-DETR (LibreDOMEDETR{s,m,l}-{aitod,visdrone})
                                                        ACADEMIC-RESEARCH-ONLY
             upstream: https://huggingface.co/RicePasteM/Dome-DETR
             revision: 530230620d1f3261a267d462989cddf204cc6e10
             Mirrored to the LibreYOLO organization under the upstream model
             card's stricter statement that the weight files are "for academic
             research purposes only". These checkpoints are NOT covered by
             LibreYOLO's MIT license and must not be presented as commercially
             cleared.

             Full context: the upstream model repo carries no license field in
             its Hugging Face metadata and no LICENSE file. Its README also
             says "This project is licensed under the Apache 2.0 license", but
             that sentence links to a missing LICENSE file. LibreYOLO's
             maintainer approved rehosting by treating the Apache statement as
             the redistribution basis and preserving the academic-only sentence
             as the controlling use restriction. This is a disclosed LibreYOLO
             interpretation, not an upstream clarification or a claim that the
             two statements are legally consistent. Every mirror uses
             `license: other`, leads with the restriction, and ships a LICENSE
             and NOTICE reproducing this context. Users remain responsible for
             reviewing the upstream terms.

             Conversion only adds LibreYOLO checkpoint metadata; state-dict
             keys and learned tensors are unchanged. Source SHA-256 values are
             recorded in the converted checkpoints and per-repository notices.
             There is no COCO checkpoint for this family: only AI-TOD-V2
             (9 classes) and VisDrone (12), hence the dataset suffix on every
             canonical filename and the absence of a bare LibreDOMEDETRs.pt.

    YOLO-NAS (LibreYOLONAS{s,m,l}, {n,s,m,l}-pose,
              {s,m,l}-obb)                             NOT REDISTRIBUTED
             upstream code: https://github.com/Deci-AI/super-gradients
                            (Apache-2.0); the rotated/OBB variant is
                            YOLO-NAS-R from pull request 2014, pinned commit
                            69141b55c1161d939939a270523a7eca5a645f72.
             upstream weights: Deci's CDN,
                            https://d2gjn4b69gu75n.cloudfront.net/models/
             The pretrained checkpoints are NOT under the repository's
             Apache-2.0 license. Detect/pose weights carry Deci's YOLO-NAS
             license and the rotated weights carry LICENSE.YOLONAS-R.md at
             the pin above: a revocable, non-transferable, non-commercial
             use grant that forbids resale, sublicensing and distribution.
             That is not a redistribution grant, so LibreYOLO mirrors none
             of them on its Hugging Face organization. Instead
             LibreYOLO("LibreYOLONASs-obb.pt") downloads
             yolo_nas_r_s_dota2.pth from Deci's CDN on demand and verifies
             it against a pinned SHA256 before unpickling. Offline
             conversion: weights/convert_yolonas_obb_weights.py.
             Complying with Deci's terms is the downstream user's
             responsibility. The DOTA2 training data is likewise not
             bundled and carries its own terms.

    YOLOv7   (LibreYOLO7b)                             MIT
             upstream: MultimediaTechLab/YOLO (Kin-Yiu Wong & Hao-Tang Tsui)
             The authors' MIT re-release, not WongKinYiu/yolov7 (GPL-3.0).
             Convert v7.pt with weights/convert_yolo7_weights.py.

    RF-DETR  (LibreRFDETR{n,s,m,l})                    Apache-2.0
             upstream: roboflow/rf-detr
             (backbone: facebookresearch/dinov2, Apache-2.0)

    RF-DETR UI (LibreRFDETRm-ui)                       MIT
             upstream: racineai/UI-DETR-1 (HF revision 0f0dda5), an
             RF-DETR-M fine-tune on Apache-2.0 base weights. Training data:
             six Roboflow Universe UI datasets, declared MIT by the authors.
             Public mirror: https://huggingface.co/LibreYOLO/LibreRFDETRm-ui

    Faster R-CNN (LibreFasterRCNN{n,s,m,l})            BSD-3-Clause (implied)
             upstream code: pytorch/vision v0.26.0 at commit
               336d36e8db990a905498c73933e35231876e28bc (BSD-3-Clause)
             The official COCO checkpoints carry no per-object license file.
             Rehosting uses BSD-3-Clause implied by the releasing project,
             not a publisher-confirmed checkpoint grant. Every weight repo
             ships the verbatim BSD text and states torchvision's warning that
             pretrained-model terms may derive from training data and users
             must determine permission for their use case.
             Public mirrors: https://huggingface.co/LibreYOLO/LibreFasterRCNN{n,s,m,l}

    RetinaNet (LibreRetinaNet{r50,r50v2})               BSD-3-Clause (implied)
             upstream code: pytorch/vision v0.26.0 at commit
               336d36e8db990a905498c73933e35231876e28bc (BSD-3-Clause)
             The official COCO checkpoints carry no per-object license file.
             Rehosting uses BSD-3-Clause implied by the releasing project,
             not a publisher-confirmed checkpoint grant. Every weight repo
             ships the verbatim BSD text and states torchvision's warning that
             pretrained-model terms may derive from training data and users
             must determine permission for their use case.
             Public mirrors: https://huggingface.co/LibreYOLO/LibreRetinaNet{r50,r50v2}
    SSD300   (LibreSSD300)                            BSD-3-Clause (implied)
             upstream code: pytorch/vision v0.26.0 at commit
               336d36e8db990a905498c73933e35231876e28bc (BSD-3-Clause)
             The official COCO checkpoint carries no per-object license file.
             Rehosting uses BSD-3-Clause implied by the releasing project and
             repeats torchvision's pretrained-model/data warning. The VGG-16
             backbone initialization traces to Oxford Visual Geometry Group
             feature weights by Karen Simonyan and Andrew Zisserman under
             CC BY 4.0; the model NOTICE preserves attribution and changes.
             Public mirror: https://huggingface.co/LibreYOLO/LibreSSD300
    Mask R-CNN (LibreMaskRCNNr50)                      BSD-3-Clause (implied)
             upstream code: pytorch/vision v0.26.0 at commit
               336d36e8db990a905498c73933e35231876e28bc (BSD-3-Clause)
             official checkpoint:
               maskrcnn_resnet50_fpn_v2_coco-73cbd019.pth
             official SHA-256:
               73cbd0190fcbe3ba339921fbce2c3a0b6bb9126c9a133c85e43a2a8e060a109e
             The checkpoint carries no per-object license file. Rehosting
             uses BSD-3-Clause implied by the releasing project, not a
             publisher-confirmed checkpoint grant. The weight repository
             ships the verbatim BSD text and torchvision pretrained-model
             caveat. Conversion adds metadata without changing learned tensors.
             Public mirror: https://huggingface.co/LibreYOLO/LibreMaskRCNNr50
    FCN      (LibreFCN{r50,r101})                      BSD-3-Clause (implied)
             upstream code: pytorch/vision v0.26.0 at commit
               336d36e8db990a905498c73933e35231876e28bc (BSD-3-Clause)
             The official COCO checkpoints carry no per-object license file.
             Rehosting uses BSD-3-Clause implied by the releasing project,
             not a publisher-confirmed checkpoint grant. Every weight repo
             ships the verbatim BSD text and states torchvision's warning that
             pretrained-model terms may derive from training data and users
             must determine permission for their use case.
             Public mirrors: https://huggingface.co/LibreYOLO/LibreFCN{r50,r101}
    CenterNet (LibreCenterNet{resdcn18,dla34})          MIT (implied)
             upstream code and official COCO checkpoints:
               xingyizhou/CenterNet at commit
               4c50fd3a46bdf63dbf2082c5cbb3458d39579e6c (MIT)
             Official checkpoint objects carry no standalone license file.
             Rehosting uses MIT implied by the releasing project and the
             maintainer-approved port brief, not a publisher-confirmed
             checkpoint-specific grant. Every mirror ships the verbatim
             CenterNet MIT license and repeats this caveat.
             resdcn18 source SHA-256:
               f9e413f91cdb235adbcb41c5c4052b8f7ff53999374048949789c29d6df18eaa
             dla34 source SHA-256:
               43bf4cc2efe00e02c1ae8484035b062a35543872d276c7dcfeb4db3e64203e4f
             Public mirrors: https://huggingface.co/LibreYOLO/LibreCenterNet{resdcn18,dla34}
    FCOS       (LibreFCOSr50)                           BSD-3-Clause (implied)
             upstream code: pytorch/vision v0.26.0 at commit
               336d36e8db990a905498c73933e35231876e28bc (BSD-3-Clause)
             The official COCO checkpoint carries no per-object license file.
             Rehosting uses BSD-3-Clause implied by the releasing project,
             not a publisher-confirmed checkpoint grant. The weight repo
             ships the verbatim BSD text and states torchvision's warning that
             pretrained-model terms may derive from training data and users
             must determine permission for their use case.
             Public mirror: https://huggingface.co/LibreYOLO/LibreFCOSr50
    AlexNet  (LibreAlexNetb-cls)                       BSD-3-Clause (implied)
             upstream code: pytorch/vision v0.26.0 at commit
               336d36e8db990a905498c73933e35231876e28bc (BSD-3-Clause)
             The official ImageNet-1K checkpoint has no per-object license
             file. Rehosting uses BSD-3-Clause implied by the releasing
             project, not a publisher-confirmed checkpoint grant. The weight
             repo ships the verbatim BSD text and states torchvision's warning
             that pretrained-model terms may derive from training data.
             Public mirror: https://huggingface.co/LibreYOLO/LibreAlexNetb-cls
    VGG      (LibreVGG{16,19,16bn,19bn}-cls)           BSD-3-Clause (implied)
             upstream code: pytorch/vision at commit
               10f68dbd78b9aa5cab9328f3b2e99cfb0b608122 (BSD-3-Clause)
             The official ImageNet-1k V1 checkpoints carry no per-object
             license file. Rehosting uses BSD-3-Clause implied by the
             releasing project, not a publisher-confirmed checkpoint grant.
             Every weight repo ships the verbatim BSD text and states
             torchvision's warning that pretrained-model terms can depend on
             training data and users must determine permission for their use
             case. Conversion leaves every learned tensor unchanged and only
             adds LibreYOLO checkpoint metadata.
             Public mirrors: https://huggingface.co/LibreYOLO/LibreVGG{16,19,16bn,19bn}-cls
    DeepLabv3 (LibreDeepLabv3{r50,r101,mv3}-sem)        BSD-3-Clause (implied)
             upstream code: pytorch/vision v0.26.0 at commit
               336d36e8db990a905498c73933e35231876e28bc (BSD-3-Clause)
             The official COCO checkpoints carry no per-object license file.
             Rehosting uses BSD-3-Clause implied by the releasing project,
             not a publisher-confirmed checkpoint grant. Every weight repo
             ships the verbatim BSD text and states torchvision's warning that
             pretrained-model terms may derive from training data and users
             must determine permission for their use case. COCO annotations
             are CC BY 4.0; source images retain individual Flickr terms.
             Public mirrors: https://huggingface.co/LibreYOLO/LibreDeepLabv3{r50,r101,mv3}-sem
    EfficientDet (LibreEfficientDet{d0,d1,d2,d3,d4})  Apache-2.0 (implied)
             upstream: rwightman/efficientdet-pytorch 0.4.1 at commit
               c6dff775a36cea0bf9b76c58e59f936411c5ce01
             Official D0-D4 COCO checkpoints come from the upstream v0.1
             GitHub release and retain every learned tensor unchanged. The
             release assets carry no separate weight-license object, so
             rehosting uses Apache-2.0 implied by the releasing project rather
             than a publisher-confirmed checkpoint-specific grant. Every
             weight repository ships the verbatim Apache-2.0 text and a NOTICE
             stating this caveat.

    Deformable DETR                                     Apache-2.0
             (LibreDeformableDETR{r50ss,r50ssdc5,r50,
              r50refine,r50twostage})
             upstream code: fundamentalvision/Deformable-DETR
             upstream weights: SenseTime/deformable-detr* on Hugging Face
             Converted without changing learned tensors; the 91-column COCO
             head is retained and mapped to contiguous COCO-80 at runtime.

    DINO-DETR (LibreDINODETR{r50,r50s5,swinl})          Apache-2.0 (implied)
             upstream code and weight release: IDEA-Research/DINO at commit
               d84a491d41898b3befd8294d1cf2614661fc0953
             Official files: checkpoint0011_4scale.pth,
               checkpoint0011_5scale.pth, and
               checkpoint0027_5scale_swin.pth from the authors' Google Drive.
             The releasing repository declares Apache-2.0, but its checkpoint
             files have no separate license file or metadata. Rehosting uses
             Apache-2.0 implied by that repository declaration, not a
             publisher-confirmed checkpoint grant. Every mirror ships the
             verbatim repository license, an explanatory NOTICE, source and
             converted SHA-256 hashes, and an exact provenance record. Learned
             tensors are unchanged; conversion only adds checkpoint metadata.
             See docs/provenance/dinodetr.md and convert with
             weights/convert_dinodetr_weights.py.

    RT-DETR  (LibreRTDETR{r18,r34,r50,r50m,r101})      Apache-2.0
             upstream: lyuwenyu/RT-DETR
             (backbone: PaddlePaddle/PaddleClas ResNet_vd, Apache-2.0)

    RT-DETRv2 OBB                                      Apache-2.0
             (LibreRTDETRv2{n,s,m,l,x}-obb)
             upstream code: RicePasteM/RiO-DETR at commit
               22d5232a4e0df6ac4bc26ed1c8aac8b4060449c7
             official DOTA 1.0 checkpoints: RicePasteM/RT-DETR-OBB at
               revision f376e9dcedfb9a47a21ac71ef61ad99f8b545698
             converted mirrors: LibreYOLO/LibreRTDETRv2{n,s,m,l,x}-obb
             Each mirror carries the upstream Apache License 2.0 and NOTICE.
             Convert a locally supplied official checkpoint with
             weights/convert_rtdetrv2_weights.py; conversion preserves learned
             tensors and adds only LibreYOLO checkpoint metadata.

    DETR     (LibreDETR{r50,r50dc5,r101,r101dc5})      Apache-2.0
             upstream: facebookresearch/detr, commit
             29901c51d7fe8712168b8d0d64351170bc0f83e0.
             Official COCO model-zoo tensors are unchanged; conversion only
             adds LibreYOLO checkpoint metadata. Every Hugging Face weight
             repository carries the upstream Apache-2.0 LICENSE and NOTICE.

    LW-DETR  (LibreLWDETR{t,s,m,l,x})                  Apache-2.0
             upstream code: Atten4Vis/LW-DETR at commit
               d5e6e6c4add2d24dafb965ced8b50163c50b9788
             upstream weights: xbsu/LW-DETR on Hugging Face, declared
             Apache-2.0 in the model-card metadata. Converted and rehosted
             at LibreYOLO/LibreLWDETR{t,s,m,l,x}; learned tensors are
             unchanged and conversion only adds LibreYOLO checkpoint
             metadata. Ported outputs are bit-exact against the official
             implementation for all five sizes; source SHA-256 hashes and
             evidence are in docs/provenance/lwdetr.md.

    EoMT     (LibreEoMTl-sem, LibreEoMTl-seg,          MIT
              LibreEoMTl-seg-1280, LibreEoMT{s,b,l}-panoptic)
             upstream: tue-mps/eomt
             (backbone: facebookresearch/dinov2, Apache-2.0)
             Panoptic checkpoints load as first-class task="panoptic"
             (133 COCO classes: 80 things + 53 stuff).

    MobileNetV4 (LibreMobileNetV4{s,m,l}-cls)          Apache-2.0
             upstream: huggingface/pytorch-image-models (timm), ImageNet-1k

    ConvNeXt (LibreConvNeXt{t,s,b}-cls)                Apache-2.0 (weights)
             upstream: timm fb_in1k; arch facebookresearch/ConvNeXt (MIT)
             (ConvNeXt V1 only; ConvNeXt V2 is listed separately below)

    ConvNeXt V2 (LibreConvNeXtV2{atto,femto,pico,n,t,b,l,h}-cls)
             Weights: CC-BY-NC-4.0 (non-commercial; redistribution permitted).
             Source: facebookresearch/ConvNeXt-V2, commit
             2553895753323c6fe0b2bf390683f5ea358a42b9; official ImageNet-1K
             224px EMA classifiers. Learned tensors are unchanged.
             Code: MIT, separate from the weight license.
             https://github.com/facebookresearch/ConvNeXt-V2/blob/2553895753323c6fe0b2bf390683f5ea358a42b9/LICENSE

    DeiT     (LibreDeiT{t,s,b}-cls)                    Apache-2.0
             upstream weights: timm *.fb_in1k at the immutable revisions and
             SHA-256 digests recorded in docs/provenance/deit.md
             architecture: facebookresearch/deit and timm (Apache-2.0)
             Only plain patch-16 224px checkpoints are mirrored; distilled,
             DeiT III, and 384px variants are NOT used.

    EfficientNetV2 (LibreEfficientNetV2{b0,b1,b2,b3}-cls)  Apache-2.0
             upstream: timm (.in1k); arch google/automl

    ResNet   (LibreResNet{18,34,50,101}-cls)           Apache-2.0
             upstream: timm (a1_in1k "ResNet Strikes Back" recipe), ImageNet-1k

    ViT      (LibreViT{ti,s,b,l}-cls)                  Apache-2.0
             upstream: timm AugReg ImageNet-21k pretraining + ImageNet-1k
             fine-tuning; metadata wrap only, learned tensors unchanged

    Swin     (LibreSwin{t,s,b,l}-cls)                  MIT
             upstream: microsoft/Swin-Transformer via timm; t/s/b ImageNet-1k,
             l ImageNet-22k pretraining to ImageNet-1k fine-tuning
             Convert with weights/convert_swin_weights.py.

    CLIP     (LibreCLIP{b32,b16}-cls)                  MIT (weights)
             upstream: mlfoundations/open_clip LAION-2B checkpoints;
             see libreyolo/models/clip/NOTICE.md for the LAION
             data-provenance note.

    SigLIP2  (LibreSigLIP2{b16,so400m}-cls)            Apache-2.0
             upstream: google/siglip2-base-patch16-256 and
             google/siglip2-so400m-patch14-384 (Hugging Face).
             Convert with weights/convert_siglip2_weights.py.

    Perception Encoder Core                            Apache-2.0
             (LibrePE{t16,s16,b16,l14,g14}-cls)
             upstream: the OpenCLIP-compatible converted repositories
             timm/PE-Core-T-16-384 @ 7fe539ed578ac49a1c2b4f946e4b0747704c825a,
             timm/PE-Core-S-16-384 @ 3249a38eb1c432ec19231c6fe1774acb6a4e4efe,
             timm/PE-Core-B-16     @ 0038414f37721c5eafc1a5e0da802d291c909de3,
             timm/PE-Core-L-14-336 @ 8eff41b3f687e50a323662c2dda5eb3588c6dd35,
             timm/PE-Core-bigG-14-448 @ 17aa0c25addfa14198fa2ff73d845a22d433432e.
             These are converted artifacts targeting the OpenCLIP-compatible
             modeling implementation, NOT unmodified official
             facebook/PE-Core-* package checkpoints. Each source model card
             declares Apache-2.0. Convert with weights/convert_pe_weights.py;
             see libreyolo/models/pe/NOTICE.md for the unresolved license
             inconsistency in the official perception_models repository.

    OMDet-Turbo (LibreOMDetTurbot)                     Apache-2.0
             upstream: omlab/omdet-turbo-swin-tiny-hf
             pinned revision:
               7fe93cecfb770c4d76cf71163956221249cab566
             LibreYOLO mirror revision:
               d569ff1da5668156f07b5859246be435804fed2d
             Model, configuration, processor, and tokenizer assets are
             byte-identical to the pinned upstream snapshot; no learned
             parameters were changed. model.safetensors SHA-256:
               439d1575d7e237ad565ed6969ea2a2dfcedf2086155e9ca3ac96cd6180a48cfd

    OV-DEIM  (LibreOVDEIM{s,m,l})                      CC BY-NC 4.0
             upstream: wleilei/OV-DEIM; converted and rehosted with
             attribution as the upstream MODEL_LICENSE permits;
             licensing confirmed by the author (wleilei/OV-DEIM#4).
             The text tower is apple/MobileCLIP-B-LT-OpenCLIP with
             tensors unmodified; the Apple license text ships in the
             weight repositories.
             Convert with weights/convert_ovdeim_weights.py.

    Depth Anything 3 (LibreDepthAnything3l-depth)      Apache-2.0
             upstream: depth-anything/DA3MONO-LARGE
             pinned revision: f465978e618db8cc79c83b8bbf24964857db1875
             Tensors are unchanged; conversion strips the outer model.
             prefix and adds LibreYOLO checkpoint metadata. Metric and
             any-view checkpoints are not hosted. CC-BY-NC-4.0
             Large/Giant/Nested checkpoints are explicitly excluded.

    MiDaS    (LibreMiDaS{s,l}-depth)                   upstream download only
             upstream: isl-org/MiDaS release assets
             midas_v21_small_256.pt SHA-256:
               70d6b9c891758c67f974a6097fb0c608c7ee67fb81ac3e5588847d5596d56fca
             dpt_large_384.pt SHA-256:
               2f21e586477d90cb9624c7eef5df7891edca49a1c4795ee2cb631fd4daa6ca69
             The releasing repository is MIT, but LibreYOLO does not bundle
             or rehost these checkpoints while commercial-use clearance of
             the mixed training datasets remains unresolved under ADR 0006.
             Runtime auto-download uses the official GitHub release URLs,
             verifies these hashes, safely loads the raw state dict, and adds
             LibreYOLO metadata locally. See docs/provenance/midas.md.

    MoGe-2  (LibreMoGe2{s,b,l}-normal)                 MIT
             upstream / pinned revision / model.pt SHA-256:
             Ruicheng/moge-2-vits-normal
               679230677b4d282c6f304189a93e98e14f085902
               79a16621928c2bf0ed04659218c55c01075e950507f40bb3332fb4c873d3e1dc
             Ruicheng/moge-2-vitb-normal
               54ad3a693e61907ea4633d13dec6ee682fa09419
               16b8110e86d5dc5a849db120ca96ef3a223fd30b0c9146d1d81db504073da5f6
             Ruicheng/moge-2-vitl-normal
               b135031bae30b5ac2ae141a0e68717795ce38340
               280741fd09bc3f403ccff9967784c2a391b52d2c0742ae3efdb21d9f90cc1a01
             All three upstream Hugging Face model cards declare license: mit.
             LibreYOLO downloads the exact pinned objects directly. Conversion
             retains the encoder, shared neck, and normal-head tensors
             unchanged, removes the unused point/mask/metric-scale heads, and
             adds LibreYOLO checkpoint metadata. Convert with
             weights/convert_moge2_weights.py.

    HRNet   (LibreHRNet{w32,w48}-pose)                 MIT implied by release
             upstream: leoxiaobin/deep-high-resolution-net.pytorch at
             6f69e4676ad8d43d0d61b64b1b9726f0c369e7b1 (MIT).
             Official source files:
               pose_hrnet_w32_256x192.pth
                 SHA-256 19bc083708bb8d873211e50d85d56344c10290c6e8b564c813fdde09645c4c1c
               pose_hrnet_w48_384x288.pth
                 SHA-256 95e0fec3194826d5e3f806ea89be68bbb84517b114c3a32b3058c56610b5ef61
             The official model zoo distributes these files from the MIT
             project but attaches no separate per-file license. LibreYOLO
             therefore describes the weight license as implied by the
             releasing project, not checkpoint-specific or independently
             publisher-confirmed. Every mirror includes the verbatim upstream
             MIT LICENSE and this limitation in NOTICE. Conversion preserves
             all learned tensors and adds only checkpoint metadata. See
             docs/provenance/hrnet.md for converted hashes and evidence.

    NAFNet   (LibreNAFNet{s,l}-restore)                no bundled weights
             upstream code: megvii-research/NAFNet (MIT)
             Some published GoPro-trained checkpoint files do not carry an
             explicit standalone weights license. Convert only weights that
             you have the right to use and redistribute. The SIDD denoising
             weights are trained on SIDD (MIT-licensed dataset).

    RealESRGAN (LibreRealESRGAN{x4,x2,x4t}-restore)    no bundled weights
             upstream code: xinntao/Real-ESRGAN (BSD-3-Clause),
             architecture lineage XPixelGroup/BasicSR (Apache-2.0)
             Released generator weights (RealESRGAN_x4plus, RealESRGAN_x2plus,
             realesr-general-x4v3) are BSD-3-Clause; mirrored as converted
             checkpoints with provenance in the model cards.

    SwinIR  (LibreSwinIR{s,m,l}-restore)               Apache-2.0
             upstream: JingyunLiang/SwinIR
             Official SwinIR-S x4 lightweight and SwinIR-M/L x4 real-world
             checkpoints. Convert with weights/convert_swinir_weights.py.

    QuickSRNet (LibreQuickSRNetm2-restore)              BSD-3-Clause
             upstream architecture/checkpoint: quic/aimet-model-zoo at
               1bd2bf5b17cdda9251437c444009b29e1a25054b
             maintained integration: qualcomm/ai-hub-models at
               16dbeb5e2805d4ada7218026de72e36878717d46
             Official source file:
               quicksrnet_medium_2x_checkpoint_float32.pth.tar
               SHA-256 a0d176b40a649e45a176c3b53f45e0237015f4f2c17b157ef5c81e38c4442a0d
             The model was trained on DIV2K. Conversion preserves all 14
             learned tensors, drops optimizer/history objects, and adds only
             LibreYOLO checkpoint metadata. The separate Hugging Face mirror
             includes the verbatim upstream BSD-3-Clause license and NOTICE.

    BiRefNet (LibreBiRefNet{t,l}-matte)                see per-size note
             upstream: ZhengPeng7/BiRefNet (code MIT)
             l (general): Hugging Face weights tagged MIT -> rehosted under
               the LibreYOLO org (LibreYOLO/LibreBiRefNetl-matte).
             t (lite): the BiRefNet_lite HF repo shows an MIT badge in its
               model card but has no explicit license metadata (no YAML
               license field, no LICENSE file). LibreYOLO does NOT rehost the
               lite weights pending an explicit license confirmation; the
               converter and a ready-to-run upload script are provided so the
               maintainer can host them once the tag is confirmed. Users may
               still convert BiRefNet_lite locally at their own discretion.

    FeyNobg (LibreFeyNobgl-matte)                      Apache-2.0
             upstream: feyninc/FeyNobg on Hugging Face (nobg library on
             GitHub), code and weights explicitly Apache-2.0 (YAML license
             tag) -> rehosted under the LibreYOLO org
             (LibreYOLO/LibreFeyNobgl-matte), including
             post-training-quantized fp16/fp8 variants
             (LibreFeyNobgl-matte-{fp16,fp8}). Upstream training data is
             not disclosed; only the released weights are redistributed.
             Convert with weights/convert_feynobg_weights.py.

    BEN2    (LibreBEN2b-matte)                         MIT
             upstream: PramaLLC/BEN2 on Hugging Face and GitHub.
             The public BEN2 Base model card declares the released weights
             MIT. LibreYOLO preserves the learned parameters unchanged and
             wraps them with checkpoint metadata. Upstream states the model
             was trained on DIS5K and Prama LLC's proprietary 22K segmentation
             dataset; LibreYOLO redistributes no training data. Convert with
             weights/convert_ben2_weights.py.

    SegFormer (LibreSegformer{b0,b1,b2,b3,b4,b5}-sem)  NON-COMMERCIAL weights
             upstream code: huggingface/transformers modeling_segformer.py
             (Apache-2.0); weights: NVIDIA's ADE20K checkpoints
             (nvidia/segformer-b0..b5-finetuned-ade-*, b5 = 640-640) under
             the NVIDIA Source Code License. That license allows
             redistribution (license copy + attribution must travel with the
             weights) but restricts USE to research or evaluation only, and
             the restriction is viral through derivative works. Hosted with a
             loud non-commercial notice printed before auto-download, like the
             VisDrone research-preview weights. NOT covered by LibreYOLO's
             permissive license. Train from scratch for unrestricted use.

    PP-LiteSeg (LibrePPLiteSeg{t50,b50,t75,b75}-sem)  NON-COMMERCIAL weights
             upstream code: Deci-AI/super-gradients ppliteseg.py / stdc.py /
             common.py / context_modules.py / dice_ce_edge_loss.py
             (Apache-2.0, commit 63de22c4), with STDC backbone lineage from
             MichaelFan01/STDC-Seg (MIT, commit 59ff37fb). Weights: the four
             released Cityscapes checkpoints, converted from
             pp_lite_{t,b}_seg{50,75}_cityscapes.pth (SHA-256 pinned in
             weights/convert_ppliteseg_weights.py). The Cityscapes license
             (https://www.cityscapes-dataset.com/license/) permits
             distributing abstract derivative models from which the dataset
             cannot be recovered, and restricts the dataset and its
             derivatives to NON-COMMERCIAL use. Hosted with the restriction
             stated in checkpoint metadata (weight_license), in the download
             notice printed before auto-download, and on the model card. NOT
             covered by LibreYOLO's permissive license; the MIT code and the
             PP-LiteSeg architecture are unaffected. Weights trained from
             scratch on your own data carry no such restriction; a fine-tune
             started from a released checkpoint inherits it.

    U-Net (LibreUNets-sem)                             Apache-2.0
             upstream: open-mmlab/mmsegmentation UNet-S5-D16 + FCN head
             (Apache-2.0, commit b040e147adfa, Copyright 2020 The
             MMSegmentation Authors). Weights: the official Cityscapes
             checkpoint
             fcn_unet_s5-d16_4x4_512x1024_160k_cityscapes_20211210_145204-6860854e.pth
             (SHA-256 pinned in weights/convert_unet_weights.py), which keeps
             the Apache-2.0 license mmsegmentation declares for it. The
             Cityscapes dataset has separate terms and is not redistributed.

    LingBot-Vision (LibreLingBotVision{s,b,l}-sem)     Apache-2.0
             upstream: robbyant/lingbot-vision (code and released backbone
             weights Apache-2.0, Copyright 2026 LingBot-Vision Contributors; backbone
             checkpoints published in the robbyant lingbot-vision Hugging
             Face collection). Each hosted checkpoint combines the unchanged
             Apache-2.0 backbone tensors with an ADE20K dense head trained
             by LibreYOLO; see each weight repo's README for the training
             data.

    PP-OCRv5 (LibrePPOCR{t,l}-ocr)                     Apache-2.0
             upstream: PaddlePaddle/PaddleOCR (code and released
             weights Apache-2.0). Converted from the official
             PP-OCRv5_{mobile,server}_{det,rec} training checkpoints;
             exact source files and SHA-256 hashes are recorded on the
             Hugging Face model cards and in
             weights/convert_ppocr_weights.py. The PP-OCRv5 recognition
             dictionary ships inside each checkpoint as charset
             metadata (Apache-2.0, attribution in
             libreyolo/models/ppocr/NOTICE).

    SAM 3    (facebook/sam3 snapshot; not rehosted)   custom SAM License
             upstream: facebook/sam3 on Hugging Face. Access is gated and
             requires accepting Meta's terms. LibreYOLO downloads the weights
             directly and does not distribute them under the MIT license.

    EdgeTAM (LibreYOLO/LibreEdgeTAM)                   Apache-2.0
             Converted from facebook/EdgeTAM revision
             14d7ecc48c656b94e5184519f698cd5386c5a2bf (edgetam.pt
             SHA-256 ed2d4850b8792c239689b043c47046ec239b6e808a3d9b6ae676c803fd8780df).
             The official EdgeTAM repository explicitly licenses both code
             and checkpoints under Apache-2.0. Conversion uses the pinned
             Apache-2.0 Transformers mapping at commit
             bd37c453544e83eb875ed3608980a1660376007a and is verified against
             yonigozlan/EdgeTAM-hf revision
             c266ce53b3fc00f0f495b583f6a116c4e57f53bb.
             Hash-pinned configuration and processor files are copied from
             that Apache-2.0-declared reference; learned tensors are converted
             independently from the official checkpoint and checked exactly.

    PicoSAM3 (LibreYOLO/LibrePicoSAM3)                 Apache-2.0
             Converted and rehosted by LibreYOLO. Upstream sources:
             upstream code commit:
               1b03949e43472953bb0021685c7fc3f5fdf48fde
             upstream weights (pietrobonazzi/picosam3) revision:
               af49e4322b6b7cf448499fee5c073d4576f59444
             Only PicoSAM3_SAM3_student_best.pt matches the advertised
             PicoSAM3 architecture. The epoch-1 files are not mirrored or
             converted because they contain the older PicoSAM2 network.
             Distilled by upstream from SAM 2.1 and SAM 3; teacher weights
             are not included or redistributed.

    SAM 3D Body (LibreSAM3DBody{d3,h}-mesh)           SAM License (restricted)
             upstream: facebookresearch/sam-3d-body (Meta Platforms, Inc.)
             The SAM License is NOT a permissive license: it carries
             field-of-use restrictions (military, nuclear, espionage and
             weapons use), a no-reverse-engineering clause, and unilateral
             amendment by Meta. It permits redistribution provided its terms
             are passed through, so the checkpoints are mirrored
             byte-identically at LibreYOLO/LibreSAM3DBodyd3-mesh and
             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: the weights remain SAM-licensed, not MIT, and are
             NOT covered by LibreYOLO's permissive license.
             No upstream code is ported; libreyolo/models/sam3dbody/ is
             LibreYOLO's own MIT wrapper around the optional upstream
             package. See THIRD_PARTY_NOTICES.txt for the full description.

    Face recognition (librefacerec-l.onnx,             per-file, see below
              librefacerec-det.onnx)
             librefacerec-l.onnx: the AuraFace-v1 embedding graph
             (glintr100.onnx from fal/AuraFace-v1 on Hugging Face,
             Apache-2.0 weights), mirrored unmodified to the LibreYOLO
             organization. Upstream training data is undisclosed; only this
             single Apache-2.0 file is mirrored, none of the other artifacts
             in the upstream repository.
             librefacerec-det.onnx: the YuNet face detector
             (face_detection_yunet_2023mar.onnx from opencv/opencv_zoo,
             MIT, Copyright (c) Shiqi Yu), mirrored unmodified.
             No code is ported from either project; both graphs are
             consumed opaquely at runtime.

    SenseNova-Vision (LibreYOLO/SenseNovaVision7b)     CC BY-NC 4.0
             upstream: OpenSenseNova/SenseNova-Vision (Apache-2.0 code,
             commit 12ccd96e32b32967a11cacb6c5bd5fe3a555fc0c). Weights are
             CC BY-NC 4.0, mirrored byte-identically with attribution at
             LibreYOLO/SenseNovaVision7b. The loader prints the
             non-commercial notice before every automatic download. NOT
             covered by LibreYOLO's permissive license. No CC BY-NC 4.0
             source is ported; see libreyolo/models/sensenova/NOTICE.

    DDColor (LibreDDColor{t,l}-restore)                Apache-2.0
             Publisher-declared artifact license at the exact pinned source
             repositories piddnad/ddcolor_paper_tiny revision
             cf9fd99c1d7472689ec7413441c1b799a51866a3 and
             piddnad/ddcolor_modelscope revision
             060f67494e31883a4b13cb27f889f3154847ada4. Converted learned
             tensors are unchanged. Both checkpoints were trained on ImageNet
             and initialized from ImageNet-22K lineage; ImageNet's data access
             agreement limits dataset use to non-commercial research and
             education. No ImageNet data is mirrored. The Artistic checkpoint
             is excluded because it also uses undisclosed private data. The
             download notice discloses this boundary; any hosted model card
             must disclose it too.

    HVI-CIDNet (LibreHVICIDNett-restore)               MIT
             Publisher-declared artifact license at
             Fediory/HVI-CIDNet-Generalization revision
             51481ef2546f870060c43eb6d6525399f5b3d2d3. Converted learned
             tensors are unchanged. The checkpoint was trained on
             LOLv2-Synthetic; the canonical LOLv2 source repository publishes
             no dataset license or explicit use terms. No training images are
             mirrored. The download notice discloses this gap; any hosted
             model card must disclose it too.

    LaMa (LibreLaMab-restore)                          Apache-2.0
             Embeds the exact OpenCV Zoo artifact
             inpainting_lama_2025jan.onnx from opencv/inpainting_lama
             revision aee6d22f0a13e5e35af1c9a1c3afd62841fc6f3f. OpenCV Zoo
             declares every file in the model directory Apache-2.0. The graph
             was trained on Places365-Challenge; its image-download terms are
             non-commercial research/education terms and do not expressly
             state trained-model terms. No Places365 image is mirrored, and
             commercial training-data clearance was not independently
             established. The download notice discloses this; any hosted
             model card must disclose it too.

    ViTMatte (LibreViTMattes-matte)                    Adobe DIM terms
             The hustvl/vitmatte-small-composition-1k checkpoint at revision
             6a58ad7646403c1df626fbd746900aec7361ea1d was trained on Adobe
             Composition-1k. Adobe's Deep Image Matting Dataset License
             Agreement says models can only be used and distributed for
             non-commercial purposes and may not be sold. Permitted
             distribution must include the exact attribution required by
             section 3. Any hosted model repository must use `license: other`
             and ship the agreement verbatim. These terms apply to the
             pretrained checkpoint, not to the separately MIT/Apache-2.0
             source port or to independently trained weights with their own
             provenance.

The LICENSE and NOTICE files on each Hugging Face model card are the
authoritative source for redistribution terms. Users converting
third-party checkpoints locally are responsible for complying with the
upstream license of those checkpoints.

TEED and DexiNed edge specialists
---------------------------------
The TEED and DexiNed architecture source is MIT-licensed. LibreYOLO bundles
the architecture ports and local checkpoint converters, but does not bundle,
mirror, or auto-download the authors' released checkpoints. Those checkpoints
were trained on BIPED, whose published dataset terms restrict use to
non-commercial purposes. Converting a locally held checkpoint does not change
its applicable terms. A checkpoint trained independently on compatibly
licensed data may be distributed under the terms established by its trainer.

V-JEPA 2 video encoders and attentive probes
-------------------------------------------
Weight licenses for this family are per artifact and are deliberately NOT
collapsed into a single family-wide claim: the two ViT-g encoders are
Apache-2.0 while the other six artifacts are MIT. Each rehosted Hugging Face
repository carries the licence of its own source artifact.

  LibreVJEPA2l256-embed.pt
    facebook/vjepa2-vitl-fpc64-256 @ b3c1679b7c34d3255ef3547f27c7b226aefab26f
    MIT
  LibreVJEPA2h256-embed.pt
    facebook/vjepa2-vith-fpc64-256 @ b5eac8703e3efdc1547fbb6ddfbeb133dc0bdee5
    MIT
  LibreVJEPA2g256-embed.pt
    facebook/vjepa2-vitg-fpc64-256 @ 875c192b7b704b87d1e1d99345769632dd5f739a
    Apache-2.0
  LibreVJEPA2g384-embed.pt
    facebook/vjepa2-vitg-fpc64-384 @ 12ca91694b230e0d4b5b0078af6f4ae1d51e933d
    Apache-2.0
  LibreVJEPA2l256-cls-ssv2.pt
    facebook/vjepa2-vitl-fpc16-256-ssv2 @ 4aa02df83918538fc21cfaf576382fa20e489a80
    MIT
  LibreVJEPA2l256-cls-diving48.pt
    facebook/vjepa2-vitl-fpc32-256-diving48 @ 71ae2a8b1ff5a297aeeaae9b5e64c7a2e5e6a633
    MIT
  LibreVJEPA2g384-cls-ssv2.pt
    facebook/vjepa2-vitg-fpc64-384-ssv2 @ 9f5fd615cb6f79065a28edcf1cc3ef25010dddfa
    MIT
  LibreVJEPA2g384-cls-diving48.pt
    facebook/vjepa2-vitg-fpc32-384-diving48 @ 0b48243375319bd8e03e3cd5560d957095429189
    MIT

The probe checkpoints were trained on Something-Something V2 and Diving48.
Those datasets are named here for provenance only. Their terms are not the
terms of these converted weights, and LibreYOLO does not mirror, bundle or
auto-download either dataset. Something-Something V2 is distributed under a
research-use agreement; Diving48's redistribution terms are not clearly
established by the selected official source. Users training on either dataset
are responsible for accepting its terms themselves.

LeVJEPA video encoder
---------------------
The converted LeVJEPA checkpoint is redistributed under Creative Commons
Attribution-NonCommercial 4.0 International. It requires attribution and may
not be used for commercial purposes. This restriction applies to the learned
weights, not LibreYOLO's independently implemented MIT-licensed runtime.

  LibreLeVJEPAl-embed.pt
    galilai-group/LeVJEPA-VideoMix-Large @ e831a0347737fcaa660b39c57d41c109de399845
    CC-BY-NC-4.0

The released checkpoint was trained on VideoMix, described by its authors as
1.8 million clips from Kinetics-710, Something-Something v2, Walking Tours,
and PE-Video. Those datasets are named for provenance only and are not
mirrored or bundled by LibreYOLO.

    Gemma 4 (LibreGemma4e2b, LibreGemma4e4b)           Apache-2.0
             mirrored snapshots of google/gemma-4-E2B-it @
             3e22461f65e89153144f8adb70e3b8c2cc9845a7 and
             google/gemma-4-E4B-it @
             ee0ef6023621cff504d758262d4e04895a5af4a2.
             Learned parameters unchanged.

    Moondream 2 (LibreMoondream2)                      Apache-2.0
             mirrored snapshot of vikhyatk/moondream2 @ 9a7d402
             (2025-06-21). Learned parameters unchanged.

    Moondream 3 (LibreMoondream3)                      BSL 1.1
             mirrored snapshot of moondream/moondream3-preview @
             5112966d1a723413b1c9a1e8bea272b72e647b35. Business Source
             License 1.1 with Additional Use Grant: redistribution is
             allowed; production use is allowed except offering the model
             to third parties as a paid hosted or embedded service that
             competes with M87 Labs's paid versions. Hosted with the
             upstream license verbatim and a banner on the card. Change
             License is Apache-2.0. NOT covered by LibreYOLO's MIT license.

    LibreGround snapshots (instruction → click)

    Florence-2-base (LibreGroundFlorence2base)         MIT
             mirrored snapshot of florence-community/Florence-2-base @
             00921df66db728a9ceb750f5eca43e5c203a2051. MIT text from
             microsoft/Florence-2-base LICENSE. Learned parameters
             unchanged.

    ShowUI-2B (LibreShowUI2b)                          MIT (weights)
             mirrored snapshot of showlab/ShowUI-2B @
             cabec4fcc48d15ffd3efe0b33ea9bc7d41509d60. The pinned
             upstream card declares license: mit. The MIT text is preserved
             verbatim from showlab/ShowUI commit
             59e059c4df62db0857cc7a7ef0b15d067a30c274. The Qwen2-VL-2B
             base is Apache-2.0; its upstream license is included in the
             mirror as LICENSE-APACHE-2.0. Those Apache obligations apply
             to the base model; they are not a re-license of the weight
             grant. Learned parameters unchanged.

    Qwen3-VL-2B (LibreGroundQwen3VL2b)                 Apache-2.0
             mirrored snapshot of Qwen/Qwen3-VL-2B-Instruct @
             89644892e4d85e24eaac8bacfd4f463576704203. Learned
             parameters unchanged.

Third-party SOURCE CODE bundled inside the LibreYOLO Python package
(separate from weights) is disclosed in the top-level NOTICE file.
Notably, the DEIMv2 family bundles DINOv3 backbone code under Meta's
non-OSI DINOv3 License Agreement; see NOTICE and
libreyolo/models/deimv2/engine/backbone/dinov3/LICENSE.md.

WildDet3D (LibreYOLO/LibreWildDet3D; byte-identical upstream mirror)
    Source: https://huggingface.co/allenai/WildDet3D revision
            161edb26239f58d8973045517d9c8b179c515f73
    License: SAM License. The model card also describes intended research
    and educational use and links Ai2's Responsible Use Guidelines.
    These weights are not MIT. The Stage 3 checkpoint is mirrored unchanged at
    LibreYOLO/LibreWildDet3D revision
    e2baa17eb2279225094a2d7e7d7620ff06dc4676 with the complete SAM License.
    Filename: wilddet3d_alldata_all_prompt_v1.0.pt
    SHA-256: f8b6a9e548f733ba62625a0d2adc4b0f4fdb6007ee11d9927f9c1027010fee57

3D-MOOD (LibreYOLO/Libre3DMOODt and LibreYOLO/Libre3DMOODb)
    Source: https://huggingface.co/RoyYang0714/3D-MOOD revision
            3d1fab552189f1a62fdb60ebb606d20625a30b90
    License: Apache-2.0, as declared by the official model repository.
    The learned parameters are unchanged; LibreYOLO renames each file for its
    canonical model-size convention and verifies the original content digest.
    Swin-T: Libre3DMOODt.pt
      Mirror revision: f1e163b249b43a8bb7b8b4ef4f446828270e35ef
      SHA-256: 699f69454625d1fb3ad7d3669d8d662d52c349c99b2aba28da1215b8e7e4a555
    Swin-B: Libre3DMOODb.pt
      Mirror revision: a4f6115d285a6983439f0e54f20c44b0ce87c261
      SHA-256: 834c976df385610bba105c7d4ea3e2d7b57dabd69dee2b61397ae398de67a677

DetAny3D (LibreYOLO/LibreDetAny3D; byte-identical upstream mirror)
    Source: https://github.com/OpenDriveLab/DetAny3D revision
            10e484be0837d80aa33cff6ed95a9be834a3f794, checkpoint published
            through the project's Google Drive folder.
    License: CC-BY-NC-4.0 applied by LibreYOLO as the strictest known term.
    OpenDriveLab declares no per-object checkpoint license; the maintainer
    approved this mirror by treating the project's Apache-2.0 statement as the
    redistribution basis for the parameters DetAny3D itself trained. That is
    LibreYOLO's disclosed interpretation, not upstream clarification. Upstream
    documents the depth branch as initialized from UniDepth v2 (CC BY-NC 4.0),
    which the repository LICENSE reproduces in full. These weights are not MIT.
    The separate runtime also includes UniDepth code under CC BY-NC 4.0.
    Filename: detany3d.pth
    Mirror revision: eeb89d1b1d37ec5a361b1ec79b26ea0620d9f4c3
    SHA-256: cd5f737ddbf3ceb64f969141672d04c0a3d785ad3b8a72668619a530840cd217

FCOS3D (LibreYOLO/LibreFCOS3D; byte-identical upstream mirror)
    Code source: https://github.com/open-mmlab/mmdetection3d (Apache-2.0).
    Checkpoint source: the OpenMMLab model zoo, finetuned R101 nuScenes
    variant published July 2021 (32.1 mAP, 39.5 NDS).
    License: no per-object checkpoint license is declared upstream. The
    maintainer approved this mirror by treating mmdetection3d's Apache-2.0
    project statement as the redistribution basis. That is LibreYOLO's
    disclosed interpretation, not upstream clarification. The weights are
    trained on nuScenes under non-commercial terms and are not MIT.
    nuScenes dataset terms: https://www.nuscenes.org/terms-of-use
    Filename: fcos3d_r101_caffe_fpn_gn-head_dcn_2x8_1x_nus-mono3d_finetune_20210717_095645-8d806dc2.pth
    Mirror revision: af69c4abfb265f63695ecbaa68b4167f0c150605
    SHA-256: 8d806dc2ecae85bc8eaba1f16dccdf03459317ca6aed7984cfda33c2a2bc33a8

Marigold V2
    Adapter source: https://huggingface.co/huawei-bayerlab/marigold-v2-0
    Revision: 6fd6d1ca246c9d2d99a4d8ac375a4eccc87178ad
    License: Apache-2.0, with a license file in the official model repository.
    Conversion preserves every inference tensor and fixed prompt tensor. Two
    training-only iREPA projector tensors are omitted explicitly.
    Frozen base: https://huggingface.co/Qwen/Qwen-Image-Edit-2509
    Revision: d3968ef930e841f4c73640fb8afa3b306a78167e
    Base license: Apache-2.0, as declared by its model card.
    The frozen base downloads separately; it is not stored in adapter .pt files.
    Verified LibreYOLO mirrors:
      LibreYOLO/LibreMarigoldV2b-depth revision 4bf4fec5b9bc64cef6ea13f424319e78f07369a5
        SHA-256: c1320b426de97cb62a2f6b972c36f8374911152694b0e1bfc58f0801445cfd81
      LibreYOLO/LibreMarigoldV2b-depth-log-stage1 revision 15d3b00db5f17b29e1db602aa1e286fc91482fe2
        SHA-256: 0d5e29a8e2badc6d1b40952b701b01a6b61116dc580bc7188140b16ab3cc599a
      LibreYOLO/LibreMarigoldV2b-depth-log-layered revision 5a7b7546935d945591b75a093623ed1c1cf22410
        SHA-256: 787cd0239aca84e9e65270c8dd66a2ca3de1a66b2367fd923b1b812275f7139b
      LibreYOLO/LibreMarigoldV2b-depth-uniform-base revision 4c2a8e0b742c515fd1ac5069f875a782aa5657eb
        SHA-256: a8cced520471685e9b22bbb01ce9055236dcc4f352bf99b7071245ae6932afec
      LibreYOLO/LibreMarigoldV2b-depth-uniform-layered revision c06fa777a491f24d30ec9683aade54cd35bf126f
        SHA-256: c91221d6a76089e275e599d188057828432cab9dd7043dc507f0853bf38289db
      LibreYOLO/LibreMarigoldV2b-depth-disparity-base revision aea778a825a32108a4083ad25050905dc6395f34
        SHA-256: 7de613be0c0fdd313891f108a8aebc7bea8b27ef0e78dac8da1931c56447b91a
      LibreYOLO/LibreMarigoldV2b-depth-disparity-layered revision 060b170064f494a0c41ae895581a0f9225b4875c
        SHA-256: 304b6f7700fcc3eeeb968d8a6627e9cdbda59493d1296b2c88a1e878cf3cf520
      LibreYOLO/LibreMarigoldV2b-normal revision 241d46b12ba01501238183bfe427cf7581a37ca6
        SHA-256: e0f6544c4f931a106efe52d8f93d736d2857e64106d8e497ab9bb792ffea96a4
      LibreYOLO/LibreMarigoldV2b-albedo revision 6047b172455c3ecca3c437b1718b06b665d6d76c
        SHA-256: 1359cc12dacab2ddbcdd81afeb637f0359fde668485915669216065155eaab71

Molmo2 (LibreMolmo2 4b / 8b / o-7b)                  Apache-2.0
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.


EdgeCrafter original COCO defaults (detection, segmentation and pose)
--------------------------------------------------------------------
License: Apache-2.0, reproduced in licenses/Apache-2.0.txt and each HF mirror.
Source: https://github.com/capsule2077/edgecrafter/releases/tag/edgecrafterv1
Assets: ecdet_{s,m,l,x}.pth, ecseg_{s,m,l,x}.pth, ecpose_{s,m,l,x}.pth.
Source license: https://github.com/capsule2077/edgecrafter/blob/b075f34de6947c615557bdb987b0cebebfb7db99/LICENSE
Copyright (c) 2026 The EdgeCrafter Authors (Longfei Liu et al., Intellindust AI Lab).

These are the earlier COCO-only checkpoints, not the Objects365 variants.
The 12 LibreYOLO HF mirrors were first populated on May 3, 2026; their latest
commits as checked on September 13, 2026 are from June 27, before the upstream
license change. Each mirror includes Apache-2.0 and a NOTICE identifying
EdgeCrafter. Example: https://huggingface.co/LibreYOLO/LibreECs
See docs/provenance/edgecrafter.md for the license history and evidence limits.


EdgeCrafter Objects365-to-COCO variants (LibreEC*-obj2coco.pt)
------------------------------------------------------------
NON-COMMERCIAL USE ONLY. Commercial use requires a separate license from
INTELLINDUST INFORMATION TECHNOLOGY (SHENZHEN) CO., LTD.

Source: https://github.com/capsule2077/edgecrafter/releases/tag/edgecrafterv1_o365
Assets: ecdet_{s,m,l,x}_o3652coco.pth, ecseg_{s,m,l,x}_o3652coco.pth,
        ecpose_{s,m,l,x}_o3652coco.pth (published August 14, 2026).
License: EdgeCrafter License, reproduced verbatim in licenses/EdgeCrafter.txt.
License source: Intellindust-AI-Lab/EdgeCrafter at
124cb0922ebe4af55c025af448312c2374210815, LICENSE.md.
Copyright (c) 2026 INTELLINDUST INFORMATION TECHNOLOGY (SHENZHEN) CO., LTD.

The asset repository retains Apache-2.0 while the main project changed to
non-commercial terms on August 20, 2026. LibreYOLO's maintainer chose to
redistribute these opt-in variants under the stricter EdgeCrafter terms.
This records the selected redistribution policy, not upstream clarification
of the conflicting notices or a claim that earlier Apache grants were revoked.
Each mirror includes the complete license, attribution, and source SHA-256.
Conversion selects the released EMA when present and adds metadata; learned
parameters are unchanged. No source code from the relicensed release is ported.

Objects365's download page limits the dataset to academic purposes, licenses
annotations under CC BY 4.0, and retains third-party image restrictions:
https://www.objects365.org/download.html
The dataset itself is not redistributed here.

Existing COCO-default checkpoints and their mirrors are unchanged. The earlier
Apache-2.0 source license is recorded at EdgeCrafter commit
744a60a864a72bf650056ffa426bc0c2aeb1920d (LICENSE), and the weight repository
still provides Apache-2.0 at b075f34de6947c615557bdb987b0cebebfb7db99 (LICENSE).
These weight terms are separate from LibreYOLO's MIT source-code license.
The August 20 replacement is commit e0656d44b80ecb97ba96bb8a804a741169431e0f;
August 24 commit 124cb0922ebe4af55c025af448312c2374210815 only renamed the file.
See docs/provenance/edgecrafter.md for links and the chronology.

GTR detection and pose (s/m/l/x)
Upstream: https://huggingface.co/Phoenix8125/GTR
Code: https://github.com/Intellindust-AI-Lab/GTR
The upstream model card explicitly declares MIT for the published weights.
COCO detection checkpoints use Objects365 pretraining. Conversion selects EMA
parameters and adds LibreYOLO metadata without changing learned tensors.
Converted checkpoints: LibreYOLO/LibreGTRs, LibreYOLO/LibreGTRm,
LibreYOLO/LibreGTRl and LibreYOLO/LibreGTRx. COCO keypoint checkpoints from
pose/ in the same upstream repository are LibreYOLO/LibreGTR{s,m,l,x}-pose.
Download revisions are pinned in libreyolo/models/gtr/model.py.

Instance segmentation (seg/gtrseg_{s,m,l,x}_coco.pth, same repository and
MIT grant, EMA tensors unchanged): LibreYOLO/LibreGTR{s,m,l,x}-seg, with
revisions pinned in LibreGTR.HF_TASK_REVISIONS.

GTR oriented boxes (s/x)
Upstream: https://huggingface.co/Phoenix8125/GTR (obb/gtrobb_{s,x}_dota.pth)
Code: https://github.com/Intellindust-AI-Lab/GTR
The upstream model card explicitly declares MIT for the published weights.
The checkpoints are fine-tuned on DOTA v1.0. Conversion selects EMA parameters and adds
LibreYOLO metadata without changing learned tensors. Converted checkpoints:
LibreYOLO/LibreGTRs-obb and LibreYOLO/LibreGTRx-obb. Download revisions are
listed in HF_TASK_REVISIONS in libreyolo/models/gtr/model.py.

GTR depth (s/m/l/x)
Upstream: https://huggingface.co/Phoenix8125/GTR (depth/gtrdepth_{s,m,l,x}.pth)
Code: https://github.com/Intellindust-AI-Lab/GTR
The upstream model card explicitly declares MIT for the published weights.
The checkpoints were pretrained on a mixed metric-depth corpus (SUN RGB-D,
DIODE, Virtual KITTI 2, KITTI, Hypersim, TartanAir, ARKitScenes and ImageNet
pseudo-labels). Conversion selects EMA
parameters and adds LibreYOLO metadata without changing learned tensors.
Converted checkpoints: LibreYOLO/LibreGTRs-depth, LibreYOLO/LibreGTRm-depth,
LibreYOLO/LibreGTRl-depth and LibreYOLO/LibreGTRx-depth.

GTR semantic segmentation (s/m/l/x, Cityscapes)
Upstream: https://huggingface.co/Phoenix8125/GTR (semseg/gtrsemseg_*_cityscapes.pth)
Code: https://github.com/Intellindust-AI-Lab/GTR
The upstream model card explicitly declares MIT for the published weights.
The checkpoints are fine-tuned on Cityscapes. Conversion (weights/convert_gtr_sem_weights.py) selects EMA parameters and adds
LibreYOLO metadata without changing learned tensors.
Converted checkpoints: LibreYOLO/LibreGTR{s,m,l,x}-sem. Download revisions are
listed in LibreGTR.HF_TASK_REVISIONS.
