Blur Studio - fotojockovic.com/blur
===================================
Blurs faces and licence plates in photos and videos, or blurs the background of a photo like a real lens
(depth-aware bokeh, tilt-shift, colour splash). Everything runs in the browser; no photo or video leaves the device.

Our own work, written for this tool (MIT licence):
  - render.js: WebGL2 lens renderer. Background separated from the subject with a weighted pull-push pyramid
    (clean plate of colour and depth behind the subject), linear-light HDR with expanded highlights, a
    scatter-as-gather disc kernel (golden-angle spiral) per pyramid level with depth ordering, round / 6-blade /
    heart / star apertures, focus from depth (circle of confusion linear in inverse depth), tilt-shift band,
    colour splash, colour decontamination of semi-transparent edges, tiled drawing for large outputs.
  - ml/gf.js: colour guided filter and fast guided-filter coefficients (matte and depth refined on the
    full-resolution photo).
  - ml/faces.js: multi-scale YuNet face finding (each scale keeps only the face sizes it is reliable at;
    overlap + containment suppression), a head fallback from hair + face-skin regions, licence-plate finding as
    a size pyramid (640 / 1024 / 1600 / 2048 px) whose large passes run as overlapping 1024 px tiles.
  - ml/mlcore.js, ml/fworker.js, ml/bworker.js: model loading with the Cache API and byte progress, WebGPU /
    WebAssembly selection and fallback, input preparation, depth normalisation.
  - track.js: video face tracks (greedy matching with a constant-velocity guess, smoothing, gap bridging,
    interpolation). facefx.js: blur / pixelate / eye bar / emoji / cover effects on a 2D canvas.
  - vworker.js: video scan and re-encode on top of Mediabunny; app.js, index.html, i18n.js, the texts and the
    translations (lang/*.js).
  - zip.js: streaming ZIP writer, copied from our FotoJock tool (MIT).
  - Model conversions done by us: IS-Net made resolution-independent and its convolution weights stored as
    int8 (one scale per output channel, DequantizeLinear in the graph); the open-lpr plate detector stored the
    same way (opset 13; same detections as the original on our test set).
    Checked against the originals with onnxruntime (IS-Net: 0.01-0.2 % of mask pixels differ).

Machine-learning models (downloaded only when needed, kept in the browser's Cache Storage):
  1. YuNet face detector (model/yunet.onnx, 0.23 MB), Shiqi Yu et al., "YuNet: A Tiny Millisecond-level Face
     Detector", Machine Intelligence Research 2023. OpenCV Zoo, face_detection_yunet_2026may.onnx.
     MIT License, Copyright (c) 2020 Shiqi Yu. Licence: model/LICENSE-YuNet.txt
     https://github.com/opencv/opencv_zoo/tree/main/models/face_detection_yunet
     Note: like almost every face detector, it was trained on the WIDER FACE dataset, whose images are
     distributed for research; the trained weights are released under MIT.
  2. Licence-plate detector: YOLOX-tiny trained by the open-lpr project (plate_yolox_tiny_640.onnx, 20.2 MB;
     here model/plates-yolox-q8.onnx, 5.2 MB, int8 weights by us). Apache License 2.0 for code and weights
     (model card: "YOLOX code and its same-repository release weights are both Apache-2.0"; Ultralytics not used).
     YOLOX architecture: Megvii-BaseDetection/YOLOX, Apache-2.0. Licence + upstream NOTICE:
     model/LICENSE-open-lpr-YOLOX.txt. https://huggingface.co/faisalthaheem/open-lpr-models
     Training data (per its NOTICE): an annotated plate corpus (github.com/faisalthaheem/open-lpr-dataset-plate-
     detection, Apache-2.0) plus unpublished community photos labelled by a Qwen3-VL teacher model.
     Downloaded only when "Blur licence plates" is on. On our test set of 70 Wikimedia Commons photos (115 plates,
     Serbian, EU, UK, US and others): 90 % of plates found at the default setting.
     Earlier versions used LPD-YuNet (OpenCV Zoo, Apache-2.0), which found about half of the European plates.
  3. MediaPipe Selfie Multiclass Segmentation 256x256 (Google) - Apache License 2.0 (model card: "Licensed under
     Apache License, Version 2.0"). Converted by us from the TFLite file to ONNX (tf2onnx) with int8 weights:
     model/selfie-mc-q8.onnx (4.4 MB), licence: model/LICENSE-MediaPipe-Selfie-Multiclass.txt. Used only for its hair and face-skin classes, to find heads that the face
     finder cannot see as a face (extreme close-ups, profiles cut by the frame).
     https://ai.google.dev/edge/mediapipe/solutions/vision/image_segmenter
  4. Depth Anything V2 Small (Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng,
     Hengshuang Zhao; 2024). Apache License 2.0 - only the Small model is Apache-2.0 (Base/Large are
     CC-BY-NC-4.0 and are NOT used). ONNX exports by onnx-community (Hugging Face), unmodified:
       model/depth-anything-v2-small-q4f16.onnx (19.1 MB, WebGPU), model/depth-anything-v2-small-int8.onnx
       (27.3 MB, WebAssembly). Licence: model/LICENSE-Depth-Anything-V2-Small.txt
     https://huggingface.co/depth-anything/Depth-Anything-V2-Small-hf, https://huggingface.co/onnx-community/depth-anything-v2-small
  5. Subject cut-out: the site's shared module /static/tools/segment/ (see its CREDITS.txt and README.md); the models
     are downloaded once per device and shared with the other fotojockovic.com tools:
     - BiRefNet (Peng Zheng, Dehong Gao, Deng-Ping Fan, Li Liu, Jorma Laaksonen, Wanli Ouyang, Nicu Sebe:
       "Bilateral Reference for High-Resolution Dichotomous Image Segmentation", CAAI AIR 2024). MIT License,
       Copyright (c) 2024 ZhengPeng. Licence: model/LICENSE-BiRefNet.txt. BiRefNet general (Swin-L, 270 MB) and BiRefNet
       lite (Swin-T, 70 MB, fallback), official ONNX exports with int8 weights by us (fp32 compute), on WebGPU. https://github.com/ZhengPeng7/BiRefNet
     - IS-Net "isnet-general-use" (Xuebin Qin, Hang Dai, Xiaobin Hu, Deng-Ping Fan, Ling Shao, Luc Van Gool:
       "Highly Accurate Dichotomous Image Segmentation", ECCV 2022). Apache License 2.0.
       Licence: model/LICENSE-IS-Net.txt. From the ONNX file distributed with rembg (MIT), int8 weights. Used
       without WebGPU. https://github.com/xuebinqin/DIS
  Not used: Ultralytics YOLO (AGPL-3.0), YOLOv9 plate models trained with the GPL-3.0 YOLOv9 code (open-image-models),
  DEIMv2 (non-commercial), InsightFace/SCRFD and RetinaFace weights from InsightFace
  (non-commercial), RMBG-1.4/2.0 (non-commercial), Robust Video Matting (GPL-3.0).

Libraries:
  - ONNX Runtime Web 1.30.0 (npm onnxruntime-web), Microsoft - MIT License. https://github.com/microsoft/onnxruntime
    ort/ort.wasm.min.js + ort/ort-wasm-glue.js + ort/ort-wasm.wasm (WebAssembly build),
    ort/ort.webgpu.min.js + ort/ort-wasm-glue-asyncify.js + ort/ort-wasm-asyncify.wasm (WebGPU build).
    The *.mjs runtime scripts were renamed to .js and the .wasm files renamed; contents unchanged
    (source-map comments removed). Licence: ort/LICENSE.txt; notices of the components compiled into it:
    ort/ThirdPartyNotices.txt
  - Mediabunny 1.61.6 (Vanilagy and contributors) - Mozilla Public License 2.0, unmodified minified bundle:
    vendor/mediabunny/mediabunny.min.js. Licence: vendor/mediabunny/LICENSE. Source:
    https://github.com/Vanilagy/mediabunny
  - Shared helpers /static/tools/common/i18n.js, rail.js and sheets.js: our own (fotojockovic.com tools).

Published methods the code builds on (ideas only; no code was copied):
  - K. He, J. Sun, X. Tang, "Guided Image Filtering", 2010/2013; K. He, J. Sun, "Fast Guided Filter", 2015.
  - D. Gustafsson, "Bokeh depth of field in a single pass", 2018 (scatter-as-gather spiral).
  - S. J. Gortler et al., "The Lumigraph", 1996 (pull-push hole filling).
  - H. Vogel, "A better way to construct the sunflower head", 1979 (golden-angle sampling).

Sample images (demo/):
  - demo/faces.jpg, faces_a/b/c.jpg: "Street Crowd" by Yann Kemper, Wikimedia Commons, CC0 1.0. Faces blurred
    and covered with this tool.
    https://commons.wikimedia.org/wiki/File:Street_Crowd.jpg
  - demo/bokeh.jpg, bokeh_a/b.jpg: "Close portrait of a young woman wearing black sunglasses in Maracaibo,
    Venezuela" by Wilfredor, Wikimedia Commons, CC0 1.0. Resized; background blurred with this tool; the
    depth and subject maps (bokeh_depth.png, bokeh_matte.png) were computed with Depth Anything V2 Small and
    BiRefNet.
    https://commons.wikimedia.org/wiki/File:Close_portrait_of_a_young_woman_wearing_black_sunglasses_in_Maracaibo,_Venezuela.jpg
  - og.jpg: made from the same pictures with this tool.

Fonts from Google Fonts (self-hosted), SIL Open Font License 1.1: Inter Tight, Instrument Serif. Emoji are drawn with
the device's own emoji font.

Self-hosted web fonts (2026-10-19): Inter Tight, Instrument Serif. SIL Open Font License 1.1. The woff2 files are the ones Google Fonts
serves (latin, latin-ext, cyrillic, greek subsets where the family has them), stored on our own server under
/static/tools/common/fonts/<family>/ with each family's licence text (OFL.txt) next to them; stylesheet
/static/tools/common/fonts/ui.css. No font request goes to Google.