Real-ESRGAN-GUI

A simple GUI wrapper for a CLI tool called realesrgan-ncnn-vulkan , a NCNN (Vulkan) implementation of Real-ESRGAN . From v1.2.0, it also integrates the GUI wrapper function of realcugan-ncnn-vulkan
, which is Real-CUGAN ‘s NCNN (Vulkan) implementation . Images such as low-resolution/low-quality illustrations and animations can be clearly and neatly enlarged (improved image quality).

I made it in about 5-6 hours in total using Flutter on Desktop (v1.0.0). Flutter
that can be easily created with a beautiful UI is the best!

install

Windows

Only Windows 10 or later 64bit OS is supported. Windows 8 and earlier and 32bit OS are not supported.

Intel Graphics, NVIDIA GPU, AMD GPU can be used for GPU like realesrgan -ncnn-vulkan / realcugan-ncnn-vulkan .

Download the latest Real-ESRGAN-GUI from the Releases page.
Real-ESRGAN-GUI-(バージョン)-windows.zipPlease download.

After the download is complete Real-ESRGAN-GUI-(バージョン)-windows.zip, unzip the file into an appropriate folder Real-ESRGAN-GUI.exeand double-click inside.
You can also try creating a shortcut on your desktop.

macOS

It supports both Intel Mac and Apple Silicon (M1, M1 Pro, M2 …etc).
I have the impression that Apple Silicon-equipped Macs generate images faster than Intel Macs.

Download the latest Real-ESRGAN-GUI from the Releases page.
Real-ESRGAN-GUI-(バージョン)-macos.zipPlease download.

Once downloaded Real-ESRGAN-GUI-(バージョン)-macos.zip, unzip the Real-ESRGAN-GUI.appfile and move it to your Applications folder.
Then Real-ESRGAN-GUI.appdouble click.

Usage

There are two modes: [File selection] for selecting files one by one, and [Folder selection] for enlarging multiple images in a folder at once.

Real-ESRGAN

It is a versatile AI image enlargement algorithm that makes images in any state reasonably high quality.
It is basically for illustration and animation, but realesrgan-x4plusif you use a model, you can also use it for live-action photos.

Compared to Real-CUGAN, lines tend to be sharper, colors are slightly different, and fine details tend to be lost.
That said, there are far fewer AI-specific artifacts, and the images look better.
In most cases, it gives a passing score, and I have the impression that the line color itself is properly maintained.

model

The model is realesr-animevideov3the fastest and most accurate (recommended).
realesrgan-x4plus-animeEdges (resolution) are more modest than , but noise is removed cleanly while maintaining the details of the original image, resulting in a natural finish. It ‘s better for
photos, but it does a decent job (although it loses some detail).realesrgan-x4plusrealesr-animevideov3

realesrgan-x4plus-animeis a realesr-animevideov3good idea to try when you’re not satisfied with how it turned out.
The result will be a realesr-animevideov3more detailed finish, but the details such as fine painting tend to be lost (although I think the difference is at a level where you can’t tell the difference unless you compare).

realesrgan-x4plusis a model that can be used not only for illustrations and animation, but also for various images. However, it is quite heavy in the environment of Intel UHD Graphics 620 (it took several minutes …).
Even when using the same image, realesrgan-x4plusthe generation of the enlarged image was completed in a few seconds in an environment equipped with NVIDIA GPU.
Since it is a general-purpose model, it can be used for both photos and illustrations, but in the case of animation, realesrgan-x4plus-animeit will be more anime-like.

Real-CUGAN

AI image enlargement algorithm specialized for illustrations and anime.
It is slower than Real-ESRGAN, realesr-animevideov3but it can enlarge the image while preserving fine detail overall.

realesr-animevideov3However, in the case of capturing animation, I have the impression that the image tends to be sharper than when the line color is darkened . Other than that, the details are certainly maintained, but it ends up being a “dark” picture, so realesr-animevideov3I think animation will give better results.

In the case of illustrations that use a lot of gradation in painting (not anime coloring), fine details and bokeh (sharpness) are maintained, so you can get better results than Real-ESRGAN, which tends to lose details. There are many.
However, I have the impression that artifacts (partial distortion of the image) are relatively more likely to occur than Real-ESRGAN, probably because it preserves details. I think it’s at a negligible level.

Another feature is that you can specify the noise reduction level.
If you set the noise reduction level to maximum, you can strongly remove JPEG noise at the cost of losing some detail.
Note that denoising levels 1 and 2 are models-seonly available when a model is specified.

Originally, it is not suitable for enlarging live-action photos at all, and without noise reduction, the image will be terrible in various ways.
However, when he dared to lower the resolution of the photo and resized it to 600px or less, he set the noise reduction level of Real-CUGAN to 3 (maximum) and enlarged it, and it was output with details like the background of the animation. .
If you manually correct the color tone, the image will look like an anime background (the point is to reduce the contrast considerably, increase the highlights and blacks, and decrease the clarity).
There may be other uses for this, such as using it for the background of an illustration.

model

The model has models-prothe highest accuracy and is recommended.
models-seAlthoughmodels-pro it is an older model, it has a wide variety of enlargement ratios and noise reduction levels.

models-nosecan only be enlarged to a magnification of 2x, but it produces a unique image with thin and sharp line edges.
However, small details are often lost.

troubleshooting

Unable to start with error “Cannot continue code execution because MSVCP140.dll was not found”

Requires installation of Visual C++ Redistributable Package 2015-2022 . After downloading
vc_redist.x64.exevc_redist.x64.exe , double-click the downloaded file to install it.

Double-click again after installation Real-ESRGAN-GUI.exeand you should be able to start it properly.

Setting the scaling factor to ‘scaled to 2x resolution’ and ‘scaled to 3x resolution’ results in catastrophic images

It is probably a bug or specification of realesrgan-ncnn-vulkan used in the backend . There is nothing we can do here…
In addition, it may be possible to expand to twice the resolution properly. I have the impression that this is likely to occur when images with high resolution such as full HD are applied to Real-ESRGAN.

It seems that Real-ESRGAN was originally developed on the premise that it will be enlarged to 4 times, so if it doesn’t work, set it to [Enlarge to 4 times resolution] and then resize it with an image editing software as appropriate.

I can’t enlarge the image with the error “failed to enlarge the image”

There are various causes, so I can’t say unconditionally, but first of all, I think that there is a possibility that the file path of the save destination is incorrect (the folder does not exist, the path specification is incorrect, etc.).

Also, if the GPU driver version is outdated, it seems that problems such as not being able to generate images or only black images are generated even if they are generated.
We recommend that you try updating your GPU driver to the latest version.