Neat Image Scaler waifu2x

http://waifu2x.udp.jp/

Found it while looking for an way to upscale the smash ultimate stock icons, shame they don’t have higher res versions in the game
insane image scaler


interesting image upscaler built on neural net
example
captain_falcon
took this tiny icon 200x200 (32.4kb) from a icon dump and…

and upscaled it to 1402x1604 (977.0kb) almost 6/7 times bigger than original!
Details were preserved fairly good. Took me 3 trys to hit the 1500x1500, which I then cropped around 400 kb and ran it again.

only downside is that the max upscale resolution is 1500x1500. Can’t wait for the day where this becomes a plugin for gimp.

Could be cool if it allowed for multiple images to be uploaded at once, towards the end I wasted more time on the captcha than physically looking and uploading the file

If I could recover my hours I should have cropped it before, as I could have gotten more details. Might try it again another time

@MazeFrame found a website that doesn’t have captcha and maxes out at 4000x4000
https://waifu2x.booru.pics/

6 Likes

Pretty cool mate, maybe change the title to better reflect the topic.

2 Likes

That’s not bad.

If only it can go bigger then 1500x1500

3 Likes

yeah thought the link would open to the page, for whatever reason HTTPS everywhere with blocking unencrypted requests and discourse doesn’t really like the website.

Shame it doesn’t have a preview of

waifu2x


Single-Image Super-Resolution for Anime-Style Art using Deep Convolutional Neural Networks. And it supports photo.

1 Like

Not bad.
Thanks for the pointer, I can definitely use this!

2 Likes


this one didn’t really turn out as nice as I liked
might try it from the artist’s source

1 Like

Maybe the algorythm works best when the original picture already is vector graphicy. As an experiment you could try to edit the image to use like 4 colours or to change contrast up or down. I’ll put this on my list of stuff to try next week.

2 Likes

seems like more objects in the frame causes more issues because feeding in this starting image
you've%20got


doesn’t have as many issues

2 Likes

makes sense :thinking:

2 Likes

Another version is this:
https://waifu2x.booru.pics/

2 Likes

def more convenient than the original one, tho doesn’t have as much noise reduction options.

Tho nothing is stopping from that owner from updating it to the highest reduction option

2 Likes

Hey, that works pretty good!

Original (375x600):

Scaled (750x1200):

3 Likes

tru test


1 Like


def has it’s perks with upscaling desktop backgrounds


1187 × 842

looks like the more detailed you can find a art the better


Taran did a guide

2 Likes

does anyone know how processing speed settings affect image quality?
image
cranking the split size makes the program take longer

using these settings

tl;dr find highest split size your system can handle, and use anything > 10 for batch size if you are using a 2070.

it’s possible that the extra settings are for people with hardware that don’t have as much compute?

If someone understands japanese and coding they can translate the readme file with how processing speed is affected by split and batch size. Would save everyone a lot of time.

Also this tool being on github also makes some interesting personal projects like

TTA makes your image less blurry but adds about 40 extra seconds in processing time.

split size Processing time
64 00:01:01.32
100 00:00:53.27
128 00:00:51.43
240 00:00:46.60
256 00:00:50.78
384 crashed
432 crashed
480 crashed
512 crashed
Output depth bits 16, Magnify + denoise, Denoise level 3, Transwidth 2560, Model 2D-Illust (UpRGB Model), Batch Size 20 System Specs Intel i7-5820K, 16 gb RAM, Gigabyte Windforce RTX 2070
Batch size Processing time
1 00:00:53.48
2 00:00:46.19
3 00:00:46.62
4 00:00:46.50
5 00:00:46.21
6 00:00:46.36
7 00:00:46.03
8 00:00:46.43
9 00:00:46.06
10 00:00:46.22
11 00:00:45.91
12 00:00:46.44
13 00:00:46.42
14 00:00:46.60
15 00:00:45.20
16 00:00:46.48
17 00:00:46.90
18 00:00:45.62
19 00:00:48.89
20 00:00:46.46

settings
Output depth bits 16
Magnify + denoise
Denoise level 3
Transwidth 2560
Model 2D-Illust (UpRGB Model)
Split Size 240

just a test data gather. This wasn’t super scientific. I didn’t really control for outside variables (was watching youtube in the background).

If I have more time, I would write a script that would make this better and I might consider writing a blog post on this.

//psuedo code of how the script would work

batch size array [x,y]
for x=1; x<21 x++
   waifu2x --settings //all custom settings
   textOutputLibrary -find "Processing time:", copy all numbers after text to variable
   save text variable y in to array[x,y]

export batchsize array to a spreadsheet with date and time and test settings as title
1 Like

@eposvox is teasing benchmarks for upscalers…