skillsafe-ai/realesrgan-x2plus

Real-ESRGAN x2plus (RRDBNet 23 blocks, 2x)

Browser-ready onnx artifacts for image-super-resolution, produced by SkillSafe's reproducible converter (models/ in skillsafe.ai) from a pinned upstream source. Every byte here is derivable from that source plus the recipe below; nothing was edited by hand.

Provenance

Upstream	https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.1/RealESRGAN_x2plus.pth
Upstream SHA-256 / commit	49fafd45f8fd7aa8d31ab2a22d14d91b536c34494a5cfe31eb5d89c2fa266abb
Recipe	recipes/realesrgan-x2plus.yaml — sha256 36c21d270546fe5fee7ad39eae6e18bbd7eaadc8d9a3faec4a270d6bb9d3abab
Toolchain	Python 3.12.13, torch 2.10.0, onnx 1.23.0, onnxruntime 1.30.0 on Darwin 25.6.0 arm64
Converted	2026-09-22T18:53:36+00:00

Files

file	class	size	SHA-256
model.onnx	registry	63.97 MB	7eb5e9eb507df603c0c04b49b23c86a362e0a01575bfbbac891d970f9c331888
model_fp16.onnx	registry	32.21 MB	382dc047b478d71200ff76e9fdb920953106a8ac63380ca7370aba357c44cc5f

registry files are parameter files served from models.skillsafe.ai once vetted; bundle files ship inside an app; registry-shared is a runtime library reused by every model of the same architecture.

Verification

ONNX Runtime (CPU) output vs the PyTorch reference on a seeded uniform input, thresholds from the recipe ({"fp32": {"max_abs": 0.0001}, "fp16": {"max_abs": 0.05, "psnr_db": 45}}):

variant	input	max abs	mean abs	PSNR
fp32	64×64	1.85e-06	3.35e-07	127.5 dB
fp32	80×96	2.03e-06	3.39e-07	127.4 dB
fp16	64×64	6.18e-04	6.55e-05	81.5 dB
fp16	80×96	6.99e-04	7.89e-05	79.8 dB

Use in the browser

import * as ort from "onnxruntime-web";
const session = await ort.InferenceSession.create("https://huggingface.co/skillsafe-ai/realesrgan-x2plus/resolve/main/model.onnx", { executionProviders: ["webgpu", "wasm"] });
const out = await session.run({ input: new ort.Tensor("float32", pixels, [1, 3, h, w]) });

Contract: input input float32 ['batch', 3, 'height', 'width'] → output output float32 ['batch', 3, 'height_x2', 'width_x2']. Opset 17.

Licence and attribution

Real-ESRGAN RealESRGAN_x2plus weights and RRDBNet architecture: Copyright (c) 2021, Xintao Wang. BSD 3-Clause License. https://github.com/xinntao/Real-ESRGAN

Licence: BSD-3-Clause — notice: https://github.com/xinntao/Real-ESRGAN/blob/a4abfb2979a7bbff3f69f58f58ae324608821e27/LICENSE. The conversion recipe and this model card are part of the SkillSafe repository and carry its licence; the weights remain under the upstream licence above.

The full manifest.json in this repo records the recipe, sources, toolchain (including the uv.lock hash) and per-file verification numbers.