Face Restore — third-party models and runtime
=============================================

Face Restore (https://face-restore.skillsafe.ai/) redistributes the following works UNMODIFIED. Each weight file is
served from SkillSafe's browser-model registry (https://models.skillsafe.ai/) and verified by SHA-256 in the browser
before use. Licence texts are in this folder.

1. YuNet face detector — face_detection_yunet_2023mar.onnx (232,589 bytes)
   Copyright (c) 2020 Shiqi Yu. MIT License (LICENSE-YuNet-MIT.txt).
   Source: https://huggingface.co/opencv/face_detection_yunet (commit 3cc26e7f1014a5ee5d74a42acee58bafc9d0a310),
   mirror of https://github.com/opencv/opencv_zoo/tree/main/models/face_detection_yunet
   SHA-256: 8f2383e4dd3cfbb4553ea8718107fc0423210dc964f9f4280604804ed2552fa4

2. GFPGAN v1.4 — gfpgan_1.4.onnx (340,299,087 bytes)
   Copyright (c) 2021 Tencent (TencentARC). Apache License 2.0 (LICENSE-GFPGAN-Apache-2.0.txt).
   Paper: Wang, Li, Zhang, Shan — "Towards Real-World Blind Face Restoration with Generative Facial Prior", CVPR 2021.
   ONNX export published by the facefusion project: https://huggingface.co/facefusion/models-3.0.0
   (commit 728b9659bd9691bf32cbf7f61af478d94b7ba81e). Upstream weights: https://github.com/TencentARC/GFPGAN
   SHA-256: accc4757b26bdb89b32b4d3500d4f79c9dff97c1dd7c7104bf9dcb95e3311385
   GFPGAN's LICENSE (shipped in full) also lists third-party source code under other terms, including the NVIDIA
   Source Code License (StyleGAN2 custom CUDA ops) and CC BY-NC-SA 4.0 (DFDNet-derived code). Face Restore ships no
   GFPGAN source code: v1.4 is GFPGAN's "clean" architecture, which runs without those custom CUDA ops, and only the
   Apache-2.0 weights (as an ONNX graph) are used here.

3. Real-ESRGAN x4plus — real_esrgan_x4_fp16.onnx (36,144,084 bytes) and real_esrgan_x4.onnx (69,539,503 bytes)
   Copyright (c) 2021 Xintao Wang. BSD 3-Clause License (LICENSE-Real-ESRGAN-BSD-3-Clause.txt).
   Paper: Wang, Xie, Dong, Shan — "Real-ESRGAN: Training Real-World Blind Super-Resolution with Pure Synthetic Data", ICCVW 2021.
   ONNX exports published by the facefusion project: https://huggingface.co/facefusion/models-3.0.0
   (commit 728b9659bd9691bf32cbf7f61af478d94b7ba81e). Upstream weights: https://github.com/xinntao/Real-ESRGAN
   SHA-256 (fp16): 223c117884ed92526ba1f89fbc09c663f15786c1049802c1a0ac8895b09a9cbc
   SHA-256 (fp32): afa7fa42eed0315a3d7b2a54e79e4e44a672d3c0889c2828c4f84c9ab91edb0b

4. onnxruntime-web 1.27.0 (ort.bundle.min.mjs, ort-wasm-simd-threaded.jsep.wasm)
   Copyright (c) Microsoft Corporation. MIT License (LICENSE-onnxruntime-MIT.txt).

The face alignment template (five FFHQ landmark positions for a 512×512 crop) follows GFPGAN's FaceRestoreHelper.
Face Restore itself is not affiliated with OpenCV, Tencent, the Real-ESRGAN authors, facefusion or Microsoft.