>>43267783TL;DR - Depends on a lot of things really. I don't use any one setting. They're different tools, with different strengths and weaknesses.
Latent - can add detail! Can add intricacy! Also seemingly completely destroys FINE detail in the first gen like fishnet fabric or other clothing textures or even cool things like "this panel on my robot waifu's leg had stripes and they're gone now" because it gets denoised during the upscale.
ESRGAN scalers - seemingly less destructive to source /detail/ but also less additive detail in the sense that it's less apt to give you "something new entirely" the way latent can and does at anything ~0.4 and higher. They do still add detail, after all, you're using the selected ESRGAN network to upscale in IMAGE space, and then denoising that with the original image in LATENT space ??? (I think. shit might as well be fucking rocket science). Think of the ESRGAN scalers -while you use them in webui- as "waifu2x. but in 2023. and with a generative image AI somehow calling some of the shots"
someone PLEASE correct me if I'm wrong on that one, kek.
I'm also lumping them all together. When you get up to ~0.6 strength, the different latent scalers differ from one another too...
She
>>43267188 was latent
bicubic antialiased - "softer" added details desired because of the style of the image at 0.6. If I know my image is being prompted sans hands - I'll go as high with 0.7 on any latent if I want "latent magic" in there, but only up to 1.5x resize. Latent itself at 2x upscaling becomes a balancing act/gacha thing at denoise >~0.5-0.6, especially if you have xformers on.
Most of the time I'm using one of the ESRGAN upscalers like lollypop - at 2x resize 0.5. That one doesn't really start to get until you get closer to 0.7. I also prefer lollypop to Anime6B at 2x and greater, but will stick to Anime6B when it's below 2x and I don't want latent fucking things up and just want {{{bigger}}}.
If you want a huge obnoxious comparison image, unfortunately with all of them at 0.7 denoising strength, lmk:
Denoising strength: 0.7, Clip skip: 2, ENSD: 31337, Hires upscale: 1.5, Hires upscaler: Latent, Script: X/Y/Z plot, X Type: Hires upscaler, X Values: "Latent, Latent (antialiased), Latent (bicubic), Latent (bicubic antialiased), Latent (nearest), Latent (nearest-exact), None, Lanczos, Nearest, 4x-AnimeSharp, 4x-UltraMix_Balanced, 4x-UltraMix_Smooth, lollypop, R-ESRGAN 4x+, R-ESRGAN 4x+ Anime6B, LDSR, ScuNET GAN, ScuNET PSNR, SwinIR 4x",
only problem is it is very hard to look at. I need to write a very simple gallery browser for individual xyz images...
And actually, looking at that wikipedia comparison of different image upscaling algorithms, I think it's neat that we can do "latent bicubic" or "latent nearest exact" but I want to see some freakshow shit like "latent xbr".
Played with xbr itself via an AviSynth script a few years ago while trying to upscale some SD - standard definition, 480p, not stable diffusion - footage and was kind of fascinated with it. I'll still use "SuperResXBR" to 4x stuff like old home videos and then Spline36 resize it back down to 720p. Looks significantly cleaner than just blowing it up to 720p with a "normal" upscaler, or letting whatever the playback device is, usually some family member's iPhone or something, handle it. 