Real-ESRGAN-x4v3 — 4× Image Super-Resolution (ONNX)

ONNX export of Real-ESRGAN's general-purpose 4× upscaler (realesr-general-x4v3 variant). At only ~5 MB it punches well above its weight on real photographs.

Re-hosted under Heliosoph for distribution stability — xinntao's GitHub releases are the authoritative source.

Credit: Wang Xintao et al. (Tencent ARC Lab).

What this repo contains

realesr-general-x4v3.onnx    # ~5 MB

A single ONNX file. The model upscales any input image to 4× its width and height with learned restoration of fine detail.

Input/output

	Spec
Input name	`input`
Input shape	dynamic — `[1, 3, H, W]`
Input dtype	float32, range `[0, 1]`
Input color order	RGB
Preprocessing	Divide by 255 (no mean/std subtraction). Tile large images to avoid OOM.
Output	`[1, 3, H4, W4]` in `[0, 1]`

How to use

import onnxruntime as ort
import numpy as np
from PIL import Image

sess = ort.InferenceSession("realesr-general-x4v3.onnx")

img = Image.open("low_res.jpg").convert("RGB")
arr = np.asarray(img, dtype=np.float32) / 255.0
arr = arr.transpose(2, 0, 1)[None, ...]  # 1x3xHxW

upscaled = sess.run(None, {"input": arr})[0][0]  # 3 x (H*4) x (W*4)
upscaled = (upscaled.transpose(1, 2, 0) * 255).clip(0, 255).astype(np.uint8)
Image.fromarray(upscaled).save("upscaled_4x.png")

For images larger than ~512×512, tile the input into overlapping patches and stitch the outputs — otherwise inference memory grows quadratically. Real-ESRGAN's reference implementation includes a tiler.

When to use which Real-ESRGAN variant

This repo ships realesr-general-x4v3 only. Other Real-ESRGAN variants exist upstream:

Variant	Best for
`realesr-general-x4v3` (this repo)	General-purpose photos, illustrations, screenshots
`RealESRGAN_x4plus_anime_6B`	Anime / cartoon-style images
`RealESRGAN_x4plus`	Photographs, more aggressive enhancement

The general-x4v3 is the safest default — it doesn't over-sharpen real photos or hallucinate on noisy input.

License

BSD-3-Clause — same as upstream. LICENSE file included.

Downloads last month: -; Downloads are not tracked for this model. How to track