QuietTake browser video eraser assets
2026-09-11

STTN
Source: https://github.com/researchmm/STTN
Revision: f39f62c5bbbe3e3eba084c487353a2c651bfdcde
License: /licenses/sttn-MIT.txt
Original sttn.pth SHA-256:
25b0c2c30042d82efd1893bd42ec726764262d94115393a1718f8d65d2a7817b
Converted to an ONNX graph with 8 RGB frames at 432 x 240. The FP32 weights
are unchanged. The FP16 variant rounds weights and activations to half
precision, retaining FP32 input/output. It is derived from the same checkpoint,
without retraining, using onnxconverter-common 1.16.0. Small pixel differences
are expected; FP32 assets remain available for devices without shader-f16.
Reproduction: services/video-erase/export_browser.py in the site repository.
FP16 reproduction: services/video-erase/export_fp16.py.

PP-OCR v4 detector
Source: https://github.com/PaddlePaddle/PaddleOCR
ONNX distribution: https://github.com/RapidAI/RapidOCR
File: rapidocr_onnxruntime/models/ch_PP-OCRv4_det_infer.onnx
License: PP-OCR-LICENSE.txt (Apache-2.0)
Weights are unchanged. Browser-side preprocessing and mask extraction are
implemented by QuietTake; no text recognition or transcription is performed.

ONNX Runtime Web
Source: https://github.com/microsoft/onnxruntime
Distribution: npm onnxruntime-web, exactly version 1.24.3
Runtime: ort-wasm-simd-threaded.asyncify.wasm and matching .mjs loader
Licenses: ORT-LICENSE.txt and ORT-ThirdPartyNotices.txt

manifest.json records SHA-256 digests and byte counts for every model/runtime
part and assembled binary. Parts are at most 20 MiB. Downloads are verified
before inference and may be cached in the visitor's browser. The browser
runs text detection, inpainting and MP4 export locally; media is not uploaded.