I finally saw the difference between SD upscaling and img2img at 2x

Been messing with stable diffusion for about 3 months now. Always just used the basic txt2img then slapped it through a standard upscaler. Got okay results. Then last week I tried taking that same low res output, running it through img2img at 0.3 denoise and 2x size. Night and day difference. The img2img version kept the structure but added actual detail instead of just smoothing out pixel mush. Took like 3 tries to get the denoise right though. Anyone else find specific workflows that actually improve the output vs just making it bigger?

2 comments

2 Comments

willow7323d ago

Three weeks ago I tried this exact same thing with a low res photo of my dog at the park. I ran the original through img2img at 0.4 denoise and 2x and it added fur texture that looked real. Then I got curious and dropped the denoise to 0.2 and it just looked like someone smeared vaseline on the screen. The sweet spot seems to be right around 0.3 for most things but I had one landscape that needed 0.45 or the clouds came out looking like mashed potatoes. I also started using the ESRGAN upscaler before img2img and that gave me even better detail without making everything look plastic. My buddy swears by dropping the CFG scale to 6 when doing this kind of thing but I keep it at 7.5 and it works fine.

james6723d ago

Fought with the same CFG scale thing for weeks and realized it's like trying to find the right water temperature in a shower. Too hot and you get burned with plastic looking garbage, too cold and nothing changes at all. Ended up keeping CFG at 7 but messing with the denoise more depending on what I'm upscaling. Found that portraits need way less denoise than landscapes because skin texture is subtle but tree bark can handle a beating. Has anyone else noticed that doing a light sharpen after all this helps more than cranking up the settings during the process?