Self-hosted portrait segmentation components
===========================================

Runtime: @mediapipe/tasks-vision 0.10.21
Publisher: Google / MediaPipe (mediapipe@google.com)
Package metadata declares Apache-2.0. Full license: LICENSE in this folder.
Upstream: https://github.com/google-ai-edge/mediapipe
Pinned package: https://registry.npmjs.org/@mediapipe/tasks-vision/-/tasks-vision-0.10.21.tgz
Package SHA-256: 343736570f433c0100ffc360dbe70f09cc502a5fc573f5cd95954692edc9cfde

The runtime .mjs and WASM/loader files are unmodified package files. The bundle's
source map is supplied under the filename named by its upstream sourceMappingURL.
Both SIMD and non-SIMD variants are included; the runtime selects one at use time.

Model: MediaPipe SelfieSegmenter (square), float16, version 1
Publisher: Google. Authors credited by the model card: Tingbo Hou,
Siargey Pisarchyk and Karthik Raveendran.
Pinned source:
https://storage.googleapis.com/mediapipe-models/image_segmenter/selfie_segmenter/float16/1/selfie_segmenter.tflite
Local path: /assets/models/selfie_segmenter.tflite
Size: 249,537 bytes
SHA-256: 191ac9529ae506ee0beefa6b2c945a172dab9d07d1e802a290a4e4038226658b
The official model card states Apache License, Version 2.0.
Full license also included as /assets/models/LICENSE.
Model card:
https://storage.googleapis.com/mediapipe-assets/Model%20Card%20MediaPipe%20Selfie%20Segmentation.pdf
Official guide:
https://developers.google.com/edge/mediapipe/solutions/vision/image_segmenter

Scope and limitations
---------------------
This person segmentation model is intended for prominent humans in portrait /
selfie images. It does not provide general object cutouts, identity recognition,
passport compliance checks or pixel-perfect masks. Fine details and difficult
lighting may be misclassified. The application exposes this limitation to users.

Application integration
-----------------------
The YourTechTool portrait engine is separate application code. It imports the
runtime and fetches the local model only after an explicit Process portrait click.
All component requests go to the same website origin. Images remain in browser
memory and are never sent to Google, an API or an upload endpoint. No analytics
or advertising scripts are enabled on the four photo pages. The passport cropper
does not load or use the AI runtime.

At first use, the selected SIMD runtime, loader, bundle and model total about
10.2 MB uncompressed, plus the small application code. Static HTTP compression
and browser caching can reduce transfers. The upload package also contains the
fallback runtime so its total size is larger.