Real-ESRGAN General x4v3 (realesr-general-x4v3)
Copyright (c) 2021, Xintao Wang. BSD 3-Clause License (LICENSE).
Model from github.com/xinntao/Real-ESRGAN; this ONNX export (real_esrgan_general_x4v3.onnx + .data, 128 × 128 input,
float) is Qualcomm AI Hub Models v0.65.0 (huggingface.co/qualcomm/Real-ESRGAN-General-x4v3, BSD-3-Clause),
downloaded 2026-10-22 from
https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/real_esrgan_general_x4v3/releases/v0.65.0/real_esrgan_general_x4v3-onnx-float.zip
SHA-256 as downloaded: onnx 09e6b675a0a18a97e1057c8dd7e984ef878a0a02ce104430da525b6241c858e7,
data 512d0ec9940c2e9d85d27f2952f12a0b77b7841dc22df4ce9f3ea458bc98f37f.
Changed 2026-10-23 (feature 115): the two files folded into one (the weights inside the .onnx file — ONNX Runtime's
CoreML on iPadOS can't find a model's weights in a file beside it) with the onnx Python package 1.23.2:
onnx.save_model(onnx.load(model, load_external_data=True), out, save_as_external_data=False). The graph and the
weights are unchanged. SHA-256 of the folded file: c6f13f8866203bc2aae3977eae6bce546b861e9e14e39682c03b17c9d6698c48.
HyriX uses it for AI upscale (feature 125), run on the user's computer by ONNX Runtime.
