BatchReady: limited evaluation beta

OPTIONAL LOCAL BACKGROUND REMOVAL
This beta offers the experimental Detail model only. The model is downloaded
from the pinned Hugging Face URL below only after explicit consent and a
high-memory acknowledgement. Processing tools come from this site. Photos and
masks are processed in a dedicated browser worker and are never uploaded.
There is no automatic download, smaller-model fallback, or paid inference API.
The earlier Quick/U2Netp model and its implementation are not included.

DETAIL MODEL: BIREFNET-LITE-512
Original BiRefNet author: ZhengPeng (ZhengPeng7), with the BiRefNet contributors.
Browser ONNX conversion: Studio Ludens V.O.F., based on ZhengPeng7/BiRefNet_lite.
Original code: https://github.com/ZhengPeng7/BiRefNet
Original weights: https://huggingface.co/ZhengPeng7/BiRefNet_lite
Converted model: https://huggingface.co/studioludens/birefnet-lite-512
Exact model card:
https://huggingface.co/studioludens/birefnet-lite-512/blob/962e67969ec6acec9c19fab76acee4f5e22e990c/README.md
Pinned revision: 962e67969ec6acec9c19fab76acee4f5e22e990c
Pinned file: onnx/model_fp16.onnx
Download URL:
https://huggingface.co/studioludens/birefnet-lite-512/resolve/962e67969ec6acec9c19fab76acee4f5e22e990c/onnx/model_fp16.onnx
Size: 98,484,532 bytes.
SHA256: eff9216bb2f9d3f023d9c2b7196845a7485739ab1f231593633e4d2344ffc516
BatchReady verifies this digest before creating an inference session.
The model is not bundled in this site's source or upload archive. Hugging Face
and its binary CDN receive normal download connection information, not photos.

MIT WEIGHTS LICENSE AND UNRESOLVED TRAINING-DATASET RIGHTS
The exact converted model declares MIT. Its complete license is included
unchanged at licenses/BIREFNET-LITE-MIT.txt, including:
Copyright (c) 2024 ZhengPeng
Copyright (c) 2026 Studio Ludens V.O.F.
The upstream author also reaffirmed the weights' MIT licensing:
https://github.com/ZhengPeng7/BiRefNet/issues/316#issuecomment-5120372022

The MIT notice does not resolve every training-dataset question. The upstream
author acknowledged the DIS5K dataset-license concern:
https://github.com/ZhengPeng7/BiRefNet/issues/306#issuecomment-4638021804
The dataset's own terms must be considered separately:
https://github.com/xuebinqin/DIS/blob/main/DIS5K-Dataset-Terms-of-Use.pdf
This limited evaluation beta proceeds with that uncertainty disclosed. It is
not a determination of full commercial clearance or a claim that every part of
the training-data provenance is unrestricted. Review those rights before a
commercial launch. This notice does not grant rights beyond the actual licenses.

PREPROCESSING AND OUTPUT LIMITS
Detail uses RGB at 512x512, /255 and ImageNet channel normalization, input_image
float32 input, output_image float32 logits, then sigmoid. Original image alpha
is multiplied by the resulting mask. Original files and their pixels remain
unchanged. Browser canvas resizing can differ from other inference runtimes.
The output can miss fine weave holes, glass, shadows, lid openings or edges.
Inspect important images and restore the original when the cutout is wrong.
The helper is preference guidance, not memory detection or quality certification.

ONNX RUNTIME WEB
Pinned official npm package: onnxruntime-web 1.30.0, Microsoft Corporation.
https://www.npmjs.com/package/onnxruntime-web/v/1.30.0
https://github.com/microsoft/onnxruntime
Runtime MIT license: licenses/ONNXRUNTIME-MIT.txt
Additional notices: licenses/ONNXRUNTIME-THIRD-PARTY.txt
The three runtime assets are unmodified from the official package:
- vendor/ort/ort.wasm.min.mjs: 50,126 bytes
  SHA256 219e6a1fc8a9938268d18efca3c91d310bd2f4a59bbd13744df5b2b7fc6cee3b
- vendor/ort/ort-wasm-simd-threaded.mjs: 24,381 bytes
  SHA256 e13f7f94fc51b4ca72b12faeb1ee95f4ace6dfbc8939bc718aabdc0a27c4299b
- vendor/ort/ort-wasm-simd-threaded.wasm: 14,239,897 bytes
  SHA256 3398c10d07d229bd91b364548e130e0e51a8e5704b88c7c083ebbeb78842dee2
Total Detail model plus runtime: 112,798,936 bytes (about 113 MB), before HTTP
compression. Progress counts the model and WASM (112,724,429 bytes); the two
small runtime modules add 74,507 bytes. Only the single-thread WASM backend is
used. The browser may reuse its ordinary HTTP cache; offline use is not promised.

Our Node WASM validation used around 2.5 GB process memory and about 10 seconds
per photo for Detail. This is not a phone benchmark. A low-memory tab may reload
or close and lose its in-page batch. No phone compatibility is certified.

SAMPLE PHOTOGRAPHS AND PREPARED CUTOUT
See samples/ATTRIBUTION.txt for the CC0 original and the explicitly prepared
Detail-model derivative used by Quick try. That entry preview runs canvas
styling, not live model inference. Live demo inference always asks permission.