Real-ESRGAN — ncnn weights
A file-for-file mirror of the ncnn model weights published in
xinntao/Real-ESRGAN release
v0.2.5.0
(realesrgan-ncnn-vulkan-20220424-ubuntu.zip). Nothing here has been retrained,
converted or quantized — these are the upstream .param/.bin pairs, byte for
byte.
They are mirrored here so that a browser application can fetch the one model it is about to run, over CORS, instead of shipping every model in a single bundle.
Files
| Model | Scale | Files | Size |
|---|---|---|---|
realesr-animevideov3 |
2× | realesr-animevideov3-x2.{param,bin} |
1.2 MB |
realesr-animevideov3 |
3× | realesr-animevideov3-x3.{param,bin} |
1.2 MB |
realesr-animevideov3 |
4× | realesr-animevideov3-x4.{param,bin} |
1.2 MB |
realesrgan-x4plus |
4× | realesrgan-x4plus.{param,bin} |
33.4 MB |
realesr-animevideov3 is an SRVGGNetCompact; realesrgan-x4plus is an RRDBNet.
Both are stored with fp16 weights, as upstream ships them.
Usage
Load a pair with any ncnn build — the blob names are data (input) and output:
ncnn::Net net;
net.opt.use_vulkan_compute = false;
net.load_param("realesrgan-x4plus.param");
net.load_model("realesrgan-x4plus.bin");
Feed the network 0..1 RGB with a 10-pixel replicated border per tile, and crop
prepadding * scale off each edge of the result — the same tiling the upstream
realesrgan-ncnn-vulkan tool uses.
License
BSD-3-Clause, as upstream. See Real-ESRGAN's LICENSE. Real-ESRGAN is by Xintao Wang, Liangbin Xie, Chao Dong and Ying Shan.
@InProceedings{wang2021realesrgan,
author = {Xintao Wang and Liangbin Xie and Chao Dong and Ying Shan},
title = {Real-ESRGAN: Training Real-World Blind Super-Resolution with Pure Synthetic Data},
booktitle = {International Conference on Computer Vision Workshops (ICCVW)},
date = {2021}
}