# Bundled local AI models

All inference is on the local CPU through Microsoft.ML.OnnxRuntime 1.24.4.
The application contains no model-download or image-upload operation.
Model hashes are checked before the first inference.

## Real-ESRGAN general x4v3

- Original project: https://github.com/xinntao/Real-ESRGAN
- Author: Xintao Wang and contributors.
- ONNX conversion distribution: https://huggingface.co/Heliosoph/realesrgan-onnx
- Download: https://huggingface.co/Heliosoph/realesrgan-onnx/resolve/main/realesr-general-x4v3.onnx
- SHA256: 09B757ACCD747D7E423C1D352B3E8F23E77CC5742D04BAE958D4EB8082B76FA4
- Licence: BSD 3-Clause; see Licenses/Real-ESRGAN-BSD-3-Clause.txt.
- RGB float NCHW input in [0,1], dynamic dimensions; native output scale 4.
- The browser uses tiled inference (192-pixel core with 32-pixel context) and up to four local CPU threads when cross-origin isolation is available. It averages the native 4x model output for 2x results.
- This restores plausible detail; it does not guarantee factual recovery of missing detail.

## U-2-Net small (u2netp)

- Original project: https://github.com/xuebinqin/U-2-Net
- Model distribution and reference preprocessing: https://github.com/danielgatis/rembg
- Download: https://github.com/danielgatis/rembg/releases/download/v0.0.0/u2netp.onnx
- SHA256: 309C8469258DDA742793DCE0EBEA8E6DD393174F89934733ECC8B14C76F4DDD8
- Upstream MD5: 8e83ca70e441ab06c318d82300c84806
- Licence: Apache 2.0; see Licenses/U-2-Net-Apache-2.0.txt.
- RGB 320x320 input, max-value scaling and ImageNet normalization; first output is a normalized subject mask.
- Erase and Restore allow manual correction. Fine hair, transparent objects and complex scenes can need refinement.

## Runtime and reference licences

ONNX Runtime: https://github.com/microsoft/onnxruntime, MIT.
rembg reference preprocessing: https://github.com/danielgatis/rembg, MIT.
Their licences are included in Licenses. Existing SkiaSharp notices are in THIRD-PARTY-NOTICES.txt.

The photo-filter implementations are original application code; no GIMP source code is bundled.

## OpenCV LaMa inpainting

- Distribution: https://huggingface.co/opencv/inpainting_lama
- File: `inpaint/inpainting_lama_2025jan.onnx` (92,591,623 bytes).
- SHA256: 7DF918AC3921D3DAF0AAE1D219776CF0DC4E4935F035AF81841B40ADCF74FDF2.
- Licence: Apache 2.0; see `inpaint/LICENSE-Apache-2.0.txt`.
- The selected rectangle becomes a local inpainting mask. Only its generated patch is blended back into the working image.
- Generated pixels are a visual reconstruction, not a recovery of the original hidden background. Results require manual review.

## Safe Check local scan (first English edition)

- Visual model: Qwen3-VL-4B-Instruct, Q4_K_M, Apache 2.0.
- GGUF distribution: https://huggingface.co/unsloth/Qwen3-VL-4B-Instruct-GGUF
- Pinned revision: 00c00da0690c4b14b5539b02c4ea5d7c9102b35e.
- Model SHA256: d4dcd426bfba75752a312b266b80fec8136fbaca13c62d93b7ac41fa67f0492b.
- F16 projector SHA256: 1b9f4e92f0fbda14d7d7b58baed86039b8a980fe503d9d6a9393f25c0028f1fc.
- CPU runtime: llama.cpp b10816, MIT, Windows x64 release.
- Runtime ZIP SHA256: b25d1044a9d6061129ce6f3439697b04996c63b9bdccf025a24a10d6d4954766.
- Models are verified at first use. Runtime is a hidden child process bound only to 127.0.0.1 on an ephemeral port, with a fresh random API key. No remote URLs are submitted to it. No photos or recognized contents are written to its logs.
- The child process exits after each scan, cancellation, or timeout. It does not install a service or use an external account.
- Text: Tesseract 5.2.0 wrapper with tessdata_fast English 4.1.0, Apache 2.0.
- Metadata: MetadataExtractor 2.9.3; QR/barcodes: ZXing.Net 0.16.11.
- OCR is English-first. Visual checks are model suggestions, not certified identity, age, ownership, rights, or safety determinations.
- The scan reports recommendations only. It does not blur, remove, upload, rename, or otherwise modify the original image.

- Standard face detector: OpenCV Zoo YuNet 2026may, MIT, pinned commit 47534e27c9851bb1128ccc0102f1145e27f23f98.
- Source: https://github.com/opencv/opencv_zoo/tree/47534e27c9851bb1128ccc0102f1145e27f23f98/models/face_detection_yunet
- face_detection_yunet.onnx SHA256: ebafce4e3c118d6554634be5c27ab333b4c047a9a8c3faf1d7cf93101c22f0f0. License: Licenses/YuNet-MIT.txt.
- Standard uses this small specialised local model, not the Advanced general vision model. It cannot establish identity, age, or consent.
- Standard OCR reads up to 4000px and bounded overlapping crops; passport/ID wording and MRZ patterns are warnings, not document authentication.