MODEL FILES - provenance, licensing and where to put them
=========================================================

NONE of these files ship with this repo; download them yourself into the ComfyUI folder named under each
section heading. Every workflow that needs one also carries a "Model:" note
on its canvas with the same download URL; this file is the licensing record.

Provenance for the images, videos and 3D assets that DO ship with the pack
(example_inputs/) lives in example_inputs/sources.txt.

Licenses below were read from the model card / LICENSE of the linked
repository. Where a model is a re-export of someone else's network, BOTH the
upstream license and the packaging license are listed - the stricter one
governs. Anything not verified says so.

Two entries need reading before use rather than after: yolo26n-seg.onnx is
AGPL-3.0 (copyleft, with a network clause), and paligemma2-3b-pt-224 is under
the Gemma Terms and is not redistributable at all.


ComfyUI/models/onnx
-------------------

face_detection_yunet_2023mar.onnx
  Title:   YuNet face detector (OpenCV Zoo)
  Creator: Shiqi Yu et al.; ONNX packaging by OpenCV Zoo
  License: MIT ("all files in this directory are licensed under MIT License")
           https://opensource.org/license/mit
  Source:  https://huggingface.co/opencv/face_detection_yunet
  Upstream: https://github.com/ShiqiYu/libfacedetection.train
           (Wu, Peng, Yu - "YuNet: a tiny millisecond-level face detector",
           Machine Intelligence Research 2023)
  Used in: 65_yunet_face_detection.json, 66_face_recognition.json

palm_detection_mediapipe_2023feb_int8bq.onnx
handpose_estimation_mediapipe_2023feb_int8bq.onnx
  Title:   MediaPipe palm detection + hand-pose estimation (OpenCV Zoo)
  Creator: Google (MediaPipe Hands); TFLite->ONNX conversion by OpenCV Zoo
  License: Apache-2.0 ("all files in this directory are licensed under
           Apache 2.0 License")
           https://www.apache.org/licenses/LICENSE-2.0
  Source:  https://huggingface.co/opencv/palm_detection_mediapipe
           https://huggingface.co/opencv/handpose_estimation_mediapipe
  Note:    converted with tensorflow-onnx + onnx-simplifier; the int8
           block-quantized variants were calibrated on FreiHAND.
  Used in: 67_handpose_estimation.json

detect_2026april.onnx + sr_2026april.onnx
  Title:   WeChat QR code detector + super-resolution models
  Creator: WeChatCV, redistributed through the OpenCV contribution repo
  License: Apache-2.0 (opencv_contribution repo license)
           https://www.apache.org/licenses/LICENSE-2.0
  Source:  https://huggingface.co/opencv/opencv_contribution
  Note:    the repo is Apache-2.0 as a whole but asks contributors to
           "refer to licenses of different models" - a per-model LICENSE may
           be stricter. Not separately verified for these two files.
  Used in: 81_qr_code_detection.json

deblurring_nafnet_2025may.onnx
  Title:   NAFNet deblurring / restoration model
  Creator: megvii-research (NAFNet); ONNX packaging by OpenCV
  License: MIT ("all files in this directory are licensed under MIT License")
           https://opensource.org/license/mit
  Source:  https://huggingface.co/opencv/deblurring_nafnet
  Upstream: https://github.com/megvii-research/NAFNet
  Used in: 72_dnn_deblur.json (the NAFNet deblur node)

mosaic-9.onnx
  Title:   ONNX fast-neural-style 'mosaic' model (Johnson et al. 2016)
  License: MIT (ONNX Model Zoo)
           https://opensource.org/license/mit
  Source:  https://huggingface.co/onnxmodelzoo/mosaic-9
  Used in: 73_dnn_style_transfer.json, 74_batch_dnn_style_transfer.json
  Note:    Swap in eudnie.onnx / candy-9.onnx / rain-princess-9.onnx /
           pointilism-9.onnx (same preprocessing) for the other styles.

east_text_detection_2026jul.onnx
  Title:   EAST text detection ONNX model (East Text Detection Demo)
  License: Apache-2.0 (OpenCV)
           https://www.apache.org/licenses/LICENSE-2.0
  Source:  https://huggingface.co/opencv/opencv_contribution
  Used in: 68_east_text_detection.json

colorization_deploy_v2_2026april.onnx
  Title:   Colorization Deploy V2 ONNX model
  Creator: OpenCV contribution (Richard Zhang / ECCV 2016)
  License: BSD-2-Clause (upstream model, (c) 2016 Richard Zhang, Phillip Isola,
           Alexei A. Efros), Apache-2.0 (OpenCV packaging)
           https://opensource.org/license/bsd-2-clause
           https://www.apache.org/licenses/LICENSE-2.0
  Source:  https://huggingface.co/opencv/opencv_contribution/tree/main/submissions/colorization
  Upstream: https://github.com/richzhang/colorization
           LICENSE: https://raw.githubusercontent.com/richzhang/colorization/master/LICENSE
           (the README's own relative link to it 404s; the raw file above is the
           text that was checked - 2 numbered conditions, no non-endorsement clause)
  Used in: 75_colorization.json

yolo26n-seg.onnx
  Title:   YOLO26-nano segmentation model
  Creator: Ultralytics (weights + exporter); conversion published through the
           OpenCV contribution repo
  License: AGPL-3.0 - the model's own README on the contribution repo states
           that "YOLO26n weights and the Ultralytics exporter are released
           under AGPL-3.0", and that any redistribution of the converted file
           is subject to those terms. The Apache-2.0 on the opencv_contribution
           repo covers the packaging only; AGPL-3.0 is the governing term.
           https://www.gnu.org/licenses/agpl-3.0.html
           https://github.com/ultralytics/ultralytics/blob/main/LICENSE
  Source:  https://huggingface.co/opencv/opencv_contribution
  Note:    THIS IS THE ONE MODEL HERE WITH A COPYLEFT NETWORK CLAUSE. Nothing
           in this pack redistributes it - the user downloads it themselves,
           which is what AGPL-3.0 leaves untouched. Serving its output over a
           network, or shipping the file inside anything, pulls the AGPL
           obligations in. The README's wording is about the converted
           yolo26n.tflite; the -seg ONNX is the same upstream weights, so the
           same terms apply.
  Used in: 70_yolo_seg.json
           (labels: example_inputs/coco-labels-2014_2017.txt, see
           example_inputs/sources.txt)

yolov4.onnx / yolov4x-mish.onnx
  Title:   YOLOv4 COCO object detector
  Creator: Bochkovskiy / Wang / Liao (darknet); ONNX packaging by OpenCV
  License: public domain (darknet's "YOLO LICENSE Version 2, July 29 2016":
           "0. Darknet is public domain. 1. Do whatever you want with it."),
           Apache-2.0 (opencv_contribution packaging)
           https://raw.githubusercontent.com/AlexeyAB/darknet/master/LICENSE
           https://www.apache.org/licenses/LICENSE-2.0
  Source:  https://huggingface.co/opencv/opencv_contribution
  Used in: 69_yolov4_detection.json
           (labels: example_inputs/coco-labels-2014_2017.txt, see
           example_inputs/sources.txt)

face_recognition_sface_2021dec.onnx   (37 MB)
  Title:   SFace face-recognition embeddings (OpenCV Zoo)
  Creator: Yaoyao Zhong (SFace loss / MobileFaceNet); ONNX conversion by
           Chengrui Wang
  License: Apache-2.0 ("all files in this directory are licensed under
           Apache 2.0 License")
           https://www.apache.org/licenses/LICENSE-2.0
  Source:  https://huggingface.co/opencv/face_recognition_sface
           https://github.com/opencv/opencv_zoo/raw/main/models/face_recognition_sface/face_recognition_sface_2021dec.onnx
  Upstream: https://github.com/zhongyy/SFace
  Used in: 66_face_recognition.json ('OpenCV SFace Embeddings')

optical_flow_estimation_raft_2023aug.onnx   (61 MB fp32)
  Title:   RAFT optical flow, ONNX export
  Creator: Zachary Teed + Jia Deng (Princeton); ONNX export by PINTO0309
  License: BSD-3-Clause (upstream RAFT), MIT (the ONNX conversion)
           https://opensource.org/license/bsd-3-clause
           https://opensource.org/license/mit
  Source:  https://huggingface.co/opencv/optical_flow_estimation_raft
  Upstream: https://github.com/princeton-vl/RAFT (arXiv 2003.12039)
  Note:    the 47 MB block-quantized ..._int8bq.onnx is in the same repo.
  Used in: 76_raft_optical_flow_dnn.json

rife_v4.9.onnx   (21 MB fp32)
  Title:   RIFE v4.9 frame interpolation, ONNX export
  Creator: Huang et al. (RIFE, ECCV 2022, arXiv 2011.06294); export by FuryTMP
  License: MIT (HF repo tag; upstream RIFE is MIT)
           https://opensource.org/license/mit
  Source:  https://huggingface.co/FuryTMP/RIFE_v4.9
  Note:    NOT the export at https://huggingface.co/FuryTMP/RIFE_fp32 -
           cv2.dnn cannot run that graph.
  Used in: 77_rife_frame_interpolation.json

squeezenet1.1-7.onnx   (~4.9 MB)
  Title:   SqueezeNet 1.1 ImageNet classifier, ONNX opset 7
  Source:  ONNX Model Zoo (validated/vision/classification/squeezenet)
  URL:     https://github.com/onnx/models/raw/main/validated/vision/classification/squeezenet/model/squeezenet1.1-7.onnx
  License: BSD-3-Clause (upstream SqueezeNet), Apache-2.0 (ONNX Model Zoo packaging)
           https://opensource.org/license/bsd-3-clause
           https://www.apache.org/licenses/LICENSE-2.0
  Used in: exercise_transfer_learning.json - as a FROZEN feature-extractor
           backbone: 'OpenCV Deep Features' taps the fire-module bottleneck
           squeezenet0_concat7 (auto-prefixed onnx_node! under the pinned
           CLASSIC dnn engine) and global-average-pools it to a 512-d
           embedding; no classifier logits, no finetuning. (The dataset itself
           is synthesized in-graph.)


ComfyUI/models/onnx/lightglue
-----------------------------

disk.onnx
  Title:   DISK feature extractor ONNX model (Tyszkiewicz et al. 2020)
  License: Apache-2.0 (LightGlue-ONNX), Apache-2.0 (DISK upstream)
           https://www.apache.org/licenses/LICENSE-2.0
  Source:  https://github.com/fabio-sim/LightGlue-ONNX/releases (v0.1.0)
  Used in: 36_feature_matching_playground.json

disk_lightglue.onnx
  Title:   LightGlue matcher ONNX model (Lindenberger et al. 2023)
  License: Apache-2.0 (LightGlue-ONNX), Apache-2.0 (LightGlue upstream)
           https://www.apache.org/licenses/LICENSE-2.0
  Source:  https://github.com/fabio-sim/LightGlue-ONNX/releases (v0.1.0)
  Used in: 36_feature_matching_playground.json


ComfyUI/models/cascades
-----------------------

haarcascade_*.xml / lbpcascade_*.xml
  Title:   Trained Viola-Jones cascades for 'OpenCV Cascade Detect'
  License: Intel/Willow Garage BSD-3-Clause (see the header of each XML)
           https://opensource.org/license/bsd-3-clause
  Source:  https://github.com/opencv/opencv/tree/4.x/data/haarcascades
           https://github.com/opencv/opencv/tree/4.x/data/lbpcascades
  Note:    OpenCV 5 DELETED the data/ folder these used to ship in
           (cv2.data.haarcascades is an empty directory in this build), so they
           must come from the 4.x branch - they still load. The 5.x URLs return
           a 404 HTML page; saved as .xml that loads as an EMPTY cascade, which
           the node reports as a config error.
  Used in: 66_face_recognition.json


ComfyUI/models/quality
----------------------

brisque_model_live.yml + brisque_range_live.yml
  Title:   BRISQUE no-reference quality model + range file, trained on LIVE
  License: Apache-2.0 (opencv_contrib)
           https://www.apache.org/licenses/LICENSE-2.0
  Source:  https://github.com/opencv/opencv_contrib
           (modules/quality/samples/)
  Used in: 86_quality_metrics.json

ComfyUI/models/llm
------------------

opencv/qwen2.5-0.5b-instruct   (folder, ~2 GB)
  Title:   Qwen2.5-0.5B-Instruct, ONNX export (fp32, with-past)
  Creator: Alibaba Qwen team; ONNX export by onnx-community
  License: Apache-2.0 (Qwen/Qwen2.5-0.5B-Instruct, the base model)
           https://www.apache.org/licenses/LICENSE-2.0
  Source:  https://huggingface.co/onnx-community/Qwen2.5-0.5B-Instruct
           (file onnx/model.onnx)
  Note:    the export repo declares no license tag of its own; the base model's
           Apache-2.0 is the governing term. The folder also needs an
           OpenCV-format config.json + the family's tokenizer.json from
           https://github.com/opencv/opencv_extra (testdata/dnn/llm/,
           Apache-2.0). Both parts of the folder matter.
  Used in: 78_llm_chat.json

opencv/paligemma2-3b-pt-224   (folder, 3-part export)
  Title:   PaliGemma2-3B pt-224 vision-language model, ONNX export
  Creator: Google; ONNX export by nklskyoy
  License: Gemma Terms of Use - NOT an open-source license and NOT
           redistributable. Gated: accept the terms on the Google model page
           before downloading.
           https://ai.google.dev/gemma/terms
           https://ai.google.dev/gemma/prohibited_use_policy
  Source:  https://huggingface.co/nklskyoy/paligemma2-3b-pt-224-onnx
           (siglip/, embedding/, gemma/ - fp32, no past inputs)
  Upstream: https://huggingface.co/google/paligemma2-3b-pt-224
  Note:    also needs the OpenCV-format config.json + the gemma2 SentencePiece
           tokenizer from https://github.com/opencv/opencv_extra
           (testdata/dnn/llm/, Apache-2.0).
  Used in: 79_vlm_caption.json


ComfyUI/models/checkpoints
--------------------------

sam3.1_multiplex_fp16.safetensors
  Title:   SAM 3.1 multiplex checkpoint, ComfyUI repack
  Creator: Meta (SAM 3.1); repacked by Comfy-Org
  License: "sam-license" (Meta's SAM License) - repo license tag. Terms not
           read here; check them before redistributing. Not an OSI license.
  Source:  https://huggingface.co/Comfy-Org/sam3.1 (checkpoints/)
  Upstream: https://huggingface.co/facebook/sam3.1
  Used in: exercise_clock_reading.json, exercise_document_perspective.json
