openpixels-app/browser-models

OpenPixels models

The ONNX files the OpenPixels web app downloads to run in the visitor's own browser: photo upscaling, denoising, face restoration, colourisation and background removal. Nothing here is new research. Every file is an existing open model, converted or repackaged so a browser can run it with onnxruntime-web, and each keeps its own upstream licence, listed below and included in full in LICENSES/.

The app downloads each file once, from a URL that names a commit of this repository rather than a branch, and checks it against the sha256 below before using it. Photos are never sent anywhere: they are processed on the visitor's device.

Files here are never replaced. A changed model is published under a new file name, so an app pinned to an earlier commit keeps working and every checksum below stays true.

Files

File	Bytes	Task	Licence
realesr_general_x4v3.onnx	4,866,422	4× upscale, strong denoise	BSD-3-Clause
realesr_general_x4v3_dn50.onnx	4,866,422	4× upscale, medium denoise	BSD-3-Clause
realesr_general_wdn_x4v3.onnx	4,866,422	4× upscale, light denoise	BSD-3-Clause
realesrgan_x4plus_w16.onnx	33,724,311	4× upscale, detailed	BSD-3-Clause
realesrgan_x4plus_anime_6b_w16.onnx	9,023,048	4× upscale, drawings and anime	BSD-3-Clause
gfpgan_1.4_w16.onnx	170,305,678	Face restoration	Apache-2.0, with the notices in GFPGAN's LICENSE (below)
ddcolor_paper_tiny_w16.onnx	115,401,645	Colourisation	Apache-2.0
u2netp.onnx	4,574,861	Background removal	Apache-2.0
modnet_photographic_portrait_matting.onnx	25,888,640	Background removal for portraits	Apache-2.0
face_detection_yunet_2023mar.onnx	232,589	Face detection	MIT

0e121299bf9b23a764d3f9e2120d0d0727d5ac5032e03a46cb78d28bcc7c6872  realesr_general_x4v3.onnx
b345bc59c38dbb78d871b5adf9519c475fe243c198f70cc9d0b97a464715d50b  realesr_general_x4v3_dn50.onnx
be536211515f088de05f1aa799a8079e92f49978b47aad6cec007f9fca62bb68  realesr_general_wdn_x4v3.onnx
29fbb9bc95708f9f63966804a389be77fe060dbdda882402bd1592a10908d364  realesrgan_x4plus_w16.onnx
26e4edb17ecd25676f58ffca1af3352cbd928bd4bc2eaec1461b7b72be0a2c41  realesrgan_x4plus_anime_6b_w16.onnx
008e67bccdcbdb611273f106dca6f88703cd02884b3ac35dbb8b3785c99a2831  gfpgan_1.4_w16.onnx
87854113d29989a6fc3c680b92b21f2cd5377a54b5babd0c5d2a53469f59244a  ddcolor_paper_tiny_w16.onnx
309c8469258dda742793dce0ebea8e6dd393174f89934733ecc8b14c76f4ddd8  u2netp.onnx
07c308cf0fc7e6e8b2065a12ed7fc07e1de8febb7dc7839d7b7f15dd66584df9  modnet_photographic_portrait_matting.onnx
8f2383e4dd3cfbb4553ea8718107fc0423210dc964f9f4280604804ed2552fa4  face_detection_yunet_2023mar.onnx

The scripts that produced every converted file, and the full provenance notes, are in the app's source repository: MODELS.md.

What _w16 means

The four *_w16.onnx files are the float32 networks with every weight of 1,024 values or more stored as float16 and followed by a Cast back to float32. Nothing else in the graph changes. onnxruntime folds the casts into ordinary float32 weights when it builds the session, so all arithmetic is still float32; what is lost is the rounding of each weight to 11 significant bits. Measured against the float32 files over ten real photographs, the outputs agree to 59–69 dB PSNR on onnxruntime's CPU backend, with no pixel more than 3 of 255 levels apart, and to 59–67 dB in Chromium on onnxruntime-web's WebGPU and WASM backends, with at most 5 of 255. The download is half the size.

Where each file comes from

Real-ESRGAN — Xintao Wang et al., https://github.com/xinntao/Real-ESRGAN. BSD 3-Clause License, Copyright (c) 2021, Xintao Wang (LICENSES/Real-ESRGAN.txt). Real-ESRGAN publishes PyTorch weights; these ONNX files are exports of them, checked against the PyTorch forward pass.

• realesr_general_x4v3.onnx — exported from realesr-general-x4v3.pth (release v0.2.5.0).
• realesr_general_wdn_x4v3.onnx — exported from realesr-general-wdn-x4v3.pth (release v0.2.5.0).
• realesr_general_x4v3_dn50.onnx — the two state dicts above interpolated 50/50 before export, which is what Real-ESRGAN's own --denoise_strength 0.5 does at run time.
• realesrgan_x4plus_w16.onnx — exported from RealESRGAN_x4plus.pth (release v0.1.0), weights stored at half size as described above.
• realesrgan_x4plus_anime_6b_w16.onnx — exported from RealESRGAN_x4plus_anime_6B.pth (release v0.2.2.4), likewise.

GFPGAN v1.4 — Tencent ARC, https://github.com/TencentARC/GFPGAN. gfpgan_1.4_w16.onnx is gfpgan_1.4.onnx as published in facefusion/models-3.0.0 (sha256 accc4757…1385), with its weights stored at half size as described above. That repository publishes the file without a model card; its name corresponds to the GFPGAN v1.4 weights, which Tencent ARC distributes as GFPGANv1.4.pth in GFPGAN release v1.3.0.

GFPGAN's LICENSE (LICENSES/GFPGAN.txt, included unchanged) reads: "GFPGAN is licensed under the Apache License Version 2.0 except for the third-party components listed below." Those components are source code in the GFPGAN repository, and two of them carry non-commercial terms: StyleGAN2 code, modified from rosinality's stylegan2-pytorch, whose upstream (NVlabs/stylegan2) is under the NVIDIA licence reproduced there, including its section 3.3 use limitation; and code "largely modified from" DFDNet, under CC BY-NC-SA 4.0. gfpgan_1.4_w16.onnx contains no source code: it is the trained GFPGAN v1.4 weights and an ONNX graph of the network they belong to, GFPGANv1Clean, whose generator follows the StyleGAN2 architecture that the listed StyleGAN2 code implements. Read the full terms before relying on this file for commercial use. GFPGAN's README separately offers a "clean" version of the architecture that does not require its customised CUDA extensions. The full text is included so the terms can be read as they are written rather than as summarised here.

DDColor — Xiaoyang Kang et al., https://github.com/piddnad/DDColor. Apache License 2.0 (LICENSES/DDColor.txt). ddcolor_paper_tiny_w16.onnx comes from the ddcolor_paper_tiny ONNX export in Qualcomm AI Hub Models release v0.63.0 (ddcolor-onnx-float.zip; model card licence apache-2.0, AI Hub Models code under BSD-3-Clause, LICENSES/qai-hub-models.txt). That release is a graph plus a separate weights file; they were merged into one file, which gives output identical to the published pair, and the weights were then stored at half size as described above.

U²-Net-p — Xuebin Qin et al., https://github.com/xuebinqin/U-2-Net. Apache License 2.0 (LICENSES/U-2-Net.txt). u2netp.onnx is the file rembg publishes in its release v0.0.0, unchanged (https://github.com/danielgatis/rembg, MIT, LICENSES/rembg.txt).

MODNet — Zhanghan Ke et al., https://github.com/ZHKKKe/MODNet, whose README states that "the code, models, and demos in this repository … are released under the Apache License 2.0" (LICENSES/MODNet.txt). modnet_photographic_portrait_matting.onnx is the official photographic portrait matting weights as redistributed, unchanged, by HivisionIDPhotos in its release pretrained-model (https://github.com/Zeyi-Lin/HivisionIDPhotos, Apache-2.0, LICENSES/HivisionIDPhotos.txt).

YuNet — Shiqi Yu et al., OpenCV Zoo, https://github.com/opencv/opencv_zoo/tree/main/models/face_detection_yunet. MIT License, Copyright (c) 2020 Shiqi Yu (LICENSES/YuNet.txt). face_detection_yunet_2023mar.onnx is the file published there, unchanged.

Changes made to the upstream works

As required by the Apache License 2.0, section 4(b): the files marked _w16 have had their weights re-stored as float16 with a cast back to float32; realesr_general_x4v3_dn50.onnx is an interpolation of two upstream state dicts; the Real-ESRGAN files are ONNX exports of PyTorch weights; and the DDColor file merges a graph with its external weights. Every other file is redistributed byte for byte as its upstream published it.