# Browser processing components and models

Files are processed locally. These runtimes and model weights are downloaded only when requested. Runtime and model version references below match the implementation; model repositories are pinned to revisions. Original source files are not sent to model hosts.

| Component | Version / revision | License | Source |
|---|---|---|---|
| Tesseract.js | 6.0.1; core 6.0.0 | Apache-2.0 | https://github.com/naptha/tesseract.js |
| Tesseract language models | projectnaptha 4.0.0; tessdata_best e12c65a915945e4c28e237a9b52bc4a8f39a0cec kor/eng/jpn | Apache-2.0 | https://github.com/tesseract-ocr/tessdata / https://github.com/tesseract-ocr/tessdata_best |
| ONNX Runtime Web | 1.30.0 | MIT (Microsoft) | https://github.com/microsoft/onnxruntime |
| Real-ESRGAN weights (x4plus, x4plus-anime-6B, general-x4v3) | ONNX exports pinned to anakhiu/realesrgan-onnx e2a8b86b, mhmtaufiq/realesrgan-onnx c524eb8b, CoderViking/realesr-general-x4v3-onnx c6a97170 — each verified tensor-by-tensor against the official checkpoints | BSD-3-Clause (Xintao Wang, Tencent ARC) | https://github.com/xinntao/Real-ESRGAN |
| Transformers.js | 4.3.0 | Apache-2.0 | https://github.com/huggingface/transformers.js |
| MODNet ONNX | fa2fa546052fba4c08921230a26cc69a333fca12 | Apache-2.0 | https://huggingface.co/Xenova/modnet / https://github.com/ZHKKKe/MODNet |
| BiRefNet lite 512 ONNX | 4a3c40c36c94093cc1e724d9ea428b8fa4b57dc7 | MIT | https://huggingface.co/studioludens/birefnet-lite-512 / https://github.com/ZhengPeng7/BiRefNet |
| Whisper tiny ONNX | 5332fcc35e32a33b86612b9a57a89be7906102b1 | Apache-2.0 (model card); MIT (Whisper code) | https://huggingface.co/Xenova/whisper-tiny / https://github.com/openai/whisper |
| gifenc | 1.0.3 | MIT (Matt DesLauriers) | https://github.com/mattdesl/gifenc |

License copies are in `licenses/` alongside this file. Libraries are loaded from their original CDN distributions without modifications. JiniTools supplies the UI, orchestration, audio resampling, masking and GIF/WebP container output.

Background extraction uses a 512px internal model for general objects, or MODNet for portraits. Models predict a mask and do not guarantee exact edge recovery. Upscaling (Real-ESRGAN) predicts details and is not recovery of original information; it runs on WebGPU when available and otherwise on single-threaded WASM. Automatic speech recognition can make transcription and timing errors. Browser and device performance varies.