[UPDATED] Comparing 4 popular upscalers in the SD Upscale

About the testing

A humble non-scientific test of comparing 4 popular upscalers in upscaling images to 2x via SD Upscale script. The data set for testing consist of 5 images that have different graphic styles and different detailed.

This test was made in the most lazy method as you can imagine, without any enhancing images via inpainting. Just upscaling after the drawing as doing many prompt-artists.

The important note: English is not my native language. Tell me if you will see any mistakes or misspellings. Will be glad to fix that.

Upscalers choose

• 4x_foolhardy_Remacri and 4x_NMKD-Siax_200k as external upscalers
  I read many good words about both. Just want to test that himself.

• SwinIR_4x and R-ESRGAN 4x+ from basic Automatic1111 bundle.
  Two basic upscalers that has good results too.

• R-ESRGAN 4x+ Anime6B
  Did not chooses for the anime upscaling because I did not see the dramatic difference between them and basic R-ESRGAN 4x+.

Upscaling settings

• Prompt: "detailed"

• Negative prompt: imputed from the original image

• Sampling steps: 40-45, depends on image

• Denoising strength: 0.1

• Clip skip: 2

• ENSD: 31337

• Scale Factor: 2

• Tile: 1600x1600

• SD upscale overlap: 192

An insane detailed image

Compare this image with 4 different upscalers

In my opinion, 4x_foolhardy_Remacri is better in this example. 4x_NMKD-Siax_200k is second.
The rest were equally bad.

Detailed anime art

Compare this image with 4 different upscalers

R-ESRGAN 4x+ was best, SwinIR_4x was second.
4x_foolhardy_Remacri is not ok but still better than 4x_NMKD-Siax_200k

Low detailed anime art

Compare this image with 4 different upscalers

This image is not good for any upscaling because the lines are so fuzzy and it has tons of noise. But we tried.

R-ESRGAN 4x+ and SwinIR_4x showed similar results, but R-ESRGAN 4x+ was a little bit better.

Detailed illustration

Compare this image with 4 different upscalers

Just regular result that can got any with art models. The small image looks good, but many details can't be upscaled correctly.

4x_foolhardy_Remacri looks a little bit better because it is not imagine details. The rest were equally.

Photo realistic image

Compare this image with 4 different upscalers

Upscaling this image before tiling and detailing was a bad idea. Do not repeat this mistake and do something with eyes before upscaling.

In my opinion, 4x_foolhardy_Remacri and 4x_NMKD-Siax_200k were similarly equal in this test. SwinIR_4x was interesting too, but not enough.

What is the best?

The target of this humble non-scientific test was to choose the best from four. But it is not simple.

4x_foolhardy_Remacri has the best average score but it fully failed in the upscaling of anime.

The rest upscalers have almost equal scores but with a nuance.
R-ESRGAN 4x+ looks good for anime but is still not enough good for art.
4x_NMKD-Siax_200k is good for detailed art and photorealistic (for photos too) but this result is not enough good for art or anime.
SwinIR_4x shows stable average results in all tests. Sounds like the multipurpose choice? Maybe.

Today I choose 4x_foolhardy_Remacri for all my tasks, just because I'm not drawing anime.
I have plans to continue researching. Adding to more the scientific moments, testing in x2/3/4, and adding the additional upscalers, of course.

How to install this additional upscalers?

1. Download that
   4x_foolhardy_Remacri
   4x_NMKD-Siax_200k
   or both here

2. Move this files to \models\ESRGAN folder and restart your SD

3. You can set this upscalers as default via configs editing.
   An example:
   "upscaler_for_img2img": "",

ui-config.json
• txt2img/Upscaler/value
• customscript/sd_upscale.py/img2img/

config.json
• upscaler_for_img2img

Credits

This test was written for small community about Stable Diffusion. Join to us!

https://t.me/capybara_paws