Sign In

SD Upscalers Comparison

SD Upscalers Comparison

Hello everybody. I have decided to make an article comparing the different upscalers for hires-fix, with denoising strength from 0.3 to 0.7, at 0.1 intervals.

I may also do model to model comparisons and others in the future. If you have any specific requests, write in the comments and I will be happy to generate and upload the results if/when I have the time.

Without further delay, here are the upscalers I will be comparing today.

At denoise strength 0.3 and 0.4:

  • 4x-AnimeSharp

  • 4x-UltraMix_Balanced

  • 4x-UltraMix_Smooth

  • 4x-UltraSharp

  • 4x_foolhardy_Remacri

  • 4x_NMKD-Superscale-SP_178000_G

  • LDSR

  • Lollypop

  • R-ESRGAN 4x+

  • R-ESRGAN 4x+ Anime6B

  • ScuNET

  • ScuNET PSNR

  • Swin2SR_RealworldSR_X4

  • SwinIR_4x

At denoise strength 0.5, 0.6 and 0.7:

  • All of the Above +

  • Latent

  • Latent (antialiased)

  • Latent (bicubic)

  • Latent (bicubic antialiased)

  • Latent (nearest)

  • Latent (nearest-exact)

I purposefully did not test the latent upscalers below 0.5, because they don't produce useable results (most of the time) under 0.45 denoising.

I tried to include as many upscalers as possible, but even so I am sure I have missed some.

If there are some upscalers that you frequently use which are not on the list, please let me know in the comments and I will try to include them in my future work.

I have prepared 4 sets of images for different styles. Let's start with illustration style.

Illustration

This is the base image we will be upscaling using hires-fix. For illustration I decided to use the Colorful model/checkpoint. if you are curious about the prompt and all the other settings, you can see them on my image post here: https://civitai.com/posts/317328

I tried to keep the prompt as similar as possible for all the different styles.

And here are the grid comparisons, below them you can also find a link for slider image comparison at denoising strength 0.5.

Grid Comparison

Slider Comparison

https://imgsli.com/MTg3MDc2

Let me mention here that I uploaded only the images with denoising 0.5 for slider comparison, because I think they will be sufficient for comparing the upscalers. If you would like me to upload all of the other images as well please write a comment and I will see if I can upload them.

Next we have Anime style.

Anime

For anime I used the MeinaMix model/checkpoint.

Again, this is the base image we will upscale with hires-fix, and you can find the prompt and all other settings on my image post here: https://civitai.com/posts/317328

Grid Comparison

Slider Comparison

https://imgsli.com/MTg3MTA2

Next up is 3D/CGI style.

3D

For 3D style I used the ReV Animated model/checkpoint.

This is the base image, prompt and settings can be found here: https://civitai.com/posts/317328

Grid Comparison

Slider Comparison

https://imgsli.com/MTg3MDc3

And the last is Realistic style.

Realistic

For realistic style I used the CyberRealistic model/checkpoint.

This is base image we are going to upscale using hires-fix. The prompt and settings used can be found here: https://civitai.com/posts/317328

Grid Comparison

Slider Comparison

https://imgsli.com/MTg3MTE0

These were all the comparisons I made for today. I hope they helped you in some way.

If you like, you can vote on the poll below as to which in your opinion are the best upscalers.

Poll

I have not used StrawPoll before, I noticed it may show ads when I created it. If you find the ads annoying drop a comment, so I won't use it in the future.

Conclusion

Making this article took up most of my time, so I haven't had the opportunity to dive into the results yet. Up to now the upscalers i used most were: UltraSharp, R-ESGGAN, LDSR, Remacri and SwinIR.

I am grateful to the community for all the resources they have created and all the help they have provided which helped me get started with Stable Diffusion.

Thank you for your time!

Buy Me A Coffee

If you liked my article you can buy me a coffee:

Ko-Fi: https://ko-fi.com/z1gamer

or crypto: https://coindrop.to/z1gamer

39

Comments