# Local object cutout — provenance and licenses

Vendored 2026-09-09. No image is sent to Google or to a third-party CDN.

- Model: Google MediaPipe MagicTouch **float32 v1**, 6,227,884 bytes. Original URL: https://storage.googleapis.com/mediapipe-models/interactive_segmenter/magic_touch/float32/1/magic_touch.tflite
- Model card: https://storage.googleapis.com/mediapipe-assets/Model%20Card%20MagicTouch.pdf explicitly states **Apache License, Version 2.0**, page 1. The 2023 card describes this 512×512×4 model. It does **not** establish that the newer 2026 split-encoder `interactive_segmenter_v2` artifact is byte-identical; that model is not included.
- Runtime: `@mediapipe/tasks-vision` **1.0.1**, Google LLC, Apache-2.0. JavaScript bundle, SIMD WASM, and no-SIMD fallback are copied unchanged from the pinned npm package. The compatible `InteractiveSegmenterLegacy` API is used for this model; migration to the stateful API requires a different model and renewed validation.
- License source: https://github.com/google-ai-edge/mediapipe/blob/v0.10.32/LICENSE — complete text beside this notice.
- Every vendored executable/model is pinned by SHA-256 in `SHA256.json`. The worker verifies the model hash before loading. New versions require an explicit change to the model, manifest, worker hash and browser smoke.
- The original MultGen worker is `cutout-worker.js`; it was written for this application and is separate from the third-party runtime.

This is interactive object segmentation, not guaranteed hair-level matting. The model card notes missed thin structures, nearby objects merging, and degraded results for small or distant subjects. Users choose a subject, adjust the confidence edge, compare a transparent preview, then explicitly apply or discard. The live canvas is never passed to the worker and Desfazer restores its prior state.

IMG.LY background-removal is no longer shipped. Its AGPL-3.0/commercial dual licensing was not treated as an implicit commercial grant. BRIA RMBG non-commercial weights and unproven rembg model provenance were not substituted. BEN2 base is a plausible MIT-licensed future alternative (https://huggingface.co/PramaLLC/BEN2), but its different resource/quality profile is not claimed as implemented here.