RealSR-NCNN-Android

中文说明

RealSR-NCNN-Android is a simple Android application that based on Waifu2x-NCNN, SRMD-NCNN, RealCUGAN-NCNN, RealSR-NCNN, & Real-ESRGAN, Anime4KCPP.
The application does not collect any private information from your device.
Download: Github Release

This repository contains 3 project:

RealSR-NCNN-Android-CLI can build programs that can be used by the console (for example, Termux) for Android.It contains 6 modules (Anime4k, RealSR, RealCUGAN, SRMD, Waifu2x and Resize)
- The RealSR program could use realsr models and real-esrgan models.
- The Resize program contains classical interpolation mode nearest bilinear bicubic and avir lancir.
RealSR-NCNN-Android-GUI can build a APK (has a GUI and easy to use). Actually it is a shell for the programs build from RealSR-NCNN-Android-CLI.
Resize-CLI just a demo like the Resize-NCNN-Android-CLI, but it not need ncnn and could build by VS.

How to use RealSR-NCNN-Android-GUI

Two ways of selecting files:

Share one or more images from other apps (e.g. Gallery) to this app
In this app, click Select Image to select an image

Tow ways of running:

chose a model, click the Run button and wait some time. The photo view will show the result when the progrem finish its work. If you like the result, you could click the Save button.
input shell command and enter. (You can input help and get more info)

input & output

Add more models to RealSR-NCNN-Android-GUI

RealSR-NCNN-Android-GUI could load extra models from sdcard automatily in ver 1.7.6. You could download more models from https://huggingface.co/tumuyan2/realsr-models .

Make a directory in sdcard.
Input the directory path to Path for custom models (RealSR/ESRGAN/Waifu2x) and save.
Download and copy models to the directory you make.
Return the main activity, then you could select the new models.

Convert pth models by yourself

Also you could convert ESRGAN pth moddls by yourself.

Download ESRGAN pytorch models from https://upscale.wiki/wiki/Model_Database and unzip it to somewhere.
Download cupscale and unzip it in your PC.
Convert pytorch models to ncnn. Open CupscaleData\bin\pth2ncnn, use pth2ncnn.exe to convert pth files to ncnn file.
Rename models, just like this:
```
models-Real-ESRGAN-AnimeSharp  // directory should have a suffix of models-Real- or models-ESRGAN-
├─x4.bin                       // models name as x[n], n is scale
├─x4.param
```
- This tool can only convert ESRGAN models, not Real-ESRGAN models. If there are Real-ESRGAN models with perfect effect that need to be converted, I can help you convert them manually.

About Real-ESRGAN

Real-ESRGAN is a Practical Algorithms for General Image Restoration.

[Paper] [Project Page] [YouTube Video] [Bilibili] [Poster] [PPT slides]
Xintao Wang, Liangbin Xie, Chao Dong, Ying Shan
Tencent ARC Lab; Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences

Note that RealESRGAN may still fail in some cases as the real-world degradations are really too complex.

About RealSR

[paper] [project] [NTIRE 2020 Challenge on Real-World Image Super-Resolution: Methods and Results]

About SRMD

[paper] [project] demo demo

About Real-CUGAN

[project]
Real-CUGAN is an AI super resolution model for anime images, trained in a million scale anime dataset, using the same architecture as Waifu2x-CUNet.

About Anime4kCPP

[[Project]](- https://github.com/TianZerL/Anime4KCPP) Anime4KCPP provides an optimized bloc97's Anime4K algorithm version 0.9, and it also provides its own CNN algorithm ACNet, it provides a variety of way to use, including preprocessing and real-time playback, it aims to be a high performance tools to process both image and video.
This project is for learning and the exploration task of algorithm course in SWJTU.

Anime4K is a simple high-quality anime upscale algorithm. The version 0.9 does not use any machine learning approaches, and can be very fast in real-time processing or pretreatment.
ACNet is a CNN based anime upscale algorithm. It aims to provide both high-quality and high-performance. HDN mode can better denoise, HDN level is from 1 to 3, higher for better denoising but may cause blur and lack of detail.

How to build RealSR-NCNN-Android-CLI

step1

https://github.com/Tencent/ncnn/releases
download ncnn-yyyymmdd-android-vulkan-shared.zip.
https://github.com/webmproject/libwebp download the source of libwebp.
https://opencv.org/releases/ download opencv-android-sdk.

step2

extract ncnn-yyyymmdd-android-vulkan-shared.zip into ../3rdparty/ncnn-android-vulkan-shared
extract the source of libwebp into ../3rdparty/libwebp
extract opencv-version-android-sdk into ../3rdparty/opencv-android-sdk

RealSR-NCNN-Android
├─3rdparty
│   ├─opencv-android-sdk
│   │   └─sdk
│   ├─libwebp
│   └─ncnn-android-vulkan-shared
│       └─arm64-v8a
├─RealSR-NCNN-Android-CLI
│   ├─Anime4k
│   ├─RealCUGAN
│   ├─Waifu2x
│   ├─RealSR
│   ├─SRMD
│   └─ReSize
└─RealSR-NCNN-Android-GUI

step3

Open this project with Android Studio, rebuild it and then you could find the program in RealSR-NCNN-Android-CLI\*\build\intermediates\cmake\release\obj\arm64-v8a or RealSR-NCNN-Android-CLI\*\build\intermediates\cmake\debug\obj\arm64-v8a
Click 3rdparty/copy_cli_build_result.bat and it could copy the build result to GUI project.

How to use RealSR-NCNN-Android-CLI

Download models

You could download assets.zip from github release page and unzip it to get models, or download models from https://github.com/tumuyan/realsr-models .

Example Command

Make sure the elf file has execute permission. Then input command

./realsr-ncnn -i input.jpg -o output.jpg

Full Usages

The usage of others program is same as realsr-ncnn.

Usage: realsr-ncnn -i infile -o outfile [options]...

  -h                   show this help
  -v                   verbose output
  -i input-path        input image path (jpg/png/webp) or directory
  -o output-path       output image path (jpg/png/webp) or directory
  -s scale             upscale ratio (4, default=4)
  -t tile-size         tile size (>=32/0=auto, default=0) can be 0,0,0 for multi-gpu
  -m model-path        realsr model path (default=models-DF2K_JPEG)
  -g gpu-id            gpu device to use (default=0) can be 0,1,2 for multi-gpu, -1 use cpu
  -j load:proc:save    thread count for load/proc/save (default=1:2:2) can be 1:2,2,2:2 for multi-gpu
  -x                   enable tta mode
  -f format            output image format (jpg/png/webp, default=ext/png)

input-path and output-path accept either file path or directory path
scale = scale level, 4 = upscale 4x
tile-size = tile size, use smaller value to reduce GPU memory usage, default selects automatically
load:proc:save = thread count for the three stages (image decoding + realsr upscaling + image encoding), using larger values may increase GPU usage and consume more GPU memory. You can tune this configuration with "4:4:4" for many small-size images, and "2:2:2" for large-size images. The default setting usually works fine for most situations. If you find that your GPU is hungry, try increasing thread count to achieve faster processing.
format = the format of the image to be output, png is better supported, however webp generally yields smaller file sizes, both are losslessly encoded

If you encounter crash or error, try to upgrade your derive

How to build RealSR-NCNN-Android-GUI

Download assets.zip from github release page, the zip file contains models & elf files. Unzip and put them to this folder, then build it with Android Studio. The direct download link for current version: https://github.com/tumuyan/RealSR-NCNN-Android/releases/download/1.9.6/assets.zip

RealSR-NCNN-Android-GUI\app\src\main\assets\
└─realsr
    │  Anime4k
    │  colors.xml
    │  delegates.xml
    │  libc++_shared.so
    │  libncnn.so
    │  libomp.so
    │  magick
    │  realcugan-ncnn
    │  realsr-ncnn
    │  resize-ncnn
    │  srmd-ncnn
    │  waifu2x-ncnn
    │  
    ├─models-nose
    │      up2x-no-denoise.bin
    │      up2x-no-denoise.param
    │      
    ├─models-pro
    │      up2x-conservative.bin
    │      up2x-conservative.param
    │      up2x-denoise3x.bin
    │      up2x-denoise3x.param
    │      up2x-no-denoise.bin
    │      up2x-no-denoise.param
    │      up3x-conservative.bin
    │      up3x-conservative.param
    │      up3x-denoise3x.bin
    │      up3x-denoise3x.param
    │      up3x-no-denoise.bin
    │      up3x-no-denoise.param
    │      
    ├─models-Real-ESRGAN
    │      x4.bin
    │      x4.param
    │      
    ├─models-Real-ESRGAN-anime
    │      x4.bin
    │      x4.param
    │      
    ├─models-Real-ESRGANv2-anime
    │      x2.bin
    │      x2.param
    │      x4.bin
    │      x4.param
    │      
    ├─models-Real-ESRGANv3-anime
    │      x2.bin
    │      x2.param
    │      x3.bin
    │      x3.param
    │      x4.bin
    │      x4.param
    │      
    ├─models-ESRGAN-Nomos8kSC
    │      x4.bin
    │      x4.param
    |
    └─models-se
           up2x-conservative.bin
           up2x-conservative.param
           up2x-denoise1x.bin
           up2x-denoise1x.param
           up2x-denoise2x.bin
           up2x-denoise2x.param
           up2x-denoise3x.bin
           up2x-denoise3x.param
           up2x-no-denoise.bin
           up2x-no-denoise.param
           up3x-conservative.bin
           up3x-conservative.param
           up3x-denoise3x.bin
           up3x-denoise3x.param
           up3x-no-denoise.bin
           up3x-no-denoise.param
           up4x-conservative.bin
           up4x-conservative.param
           up4x-denoise3x.bin
           up4x-denoise3x.param
           up4x-no-denoise.bin
           up4x-no-denoise.param

Acknowledgement

original super-resolution projects

ncnn projects and models

Most of the C code is copied from Nihui, cause of the directory structure had to be adjusted, the original git was broken

Others Open-Source Code Used

https://github.com/Tencent/ncnn for fast neural network inference on ALL PLATFORMS
https://github.com/nothings/stb for decoding and encoding image on Linux / MacOS
https://github.com/tronkko/dirent for listing files in directory on Windows
https://github.com/webmproject/libwebp for encoding and decoding Webp images on ALL PLATFORMS
https://github.com/avaneev/avir AVIR image resizing algorithm designed by Aleksey Vaneev
https://github.com/ImageMagick/ImageMagick6 Use ImageMagick® to resize/convert images.
https://github.com/MolotovCherry/Android-ImageMagick7 The repository has been archived, I use this fork https://github.com/Miseryset/Android-ImageMagick7

Others packaged models

Real-ESRGAN model Nomos8kSC trained by Phhofm.

tumuyan / RealSR-NCNN-Android

readme