>>5142

...so magical! xD

>>5139

>Hope you get well soon, merry late christmas!

thanks! :)

>Maybe it would be easier for your PC if they weren't so high resolution (...), if that's something you set? I don't know much about technical details of these so I'm probably wrong

No you're not wrong. It would be faster if I'd work with smaller resolutions, yes. Pony Diffusion and its mixxes (and all other models that are based on SDXL) are trained to be best used with resolutions at about 1024x1024 pixels. So for other aspect ratios you'll ideally end up with one value above and one value below that for your initial generations. While it can generate smaller pics much faster the quality diminishes further and the subsequent steps to fix all the errors in the pics often becomes even more time consuming in the end, especially with char-Loras and guro-Loras (in my humble experience). The multidiffusion step upscales and doubles the resolution from the initial gen, the time consuming part here is that I let it run at a denoise strength of 0.33 at first to get somewhere between 6-20 very similar upscaled/enhanced versions of that image. (the number of pics is depending on how good, large, detailed etc. the initial gen is, that I'm trying to refine. The denoise value defines how much the image gets scrambled (introduction of noise) before it then gets restored, upscaled and so on by multidiffusion). Then I pick the one image - seed out of that run that I think is overall looking the best and let multidiffusion run on just that seed again with a script active, that will repeat the multidiffusion with different denoise values in predetermined steps. Usually I'll use denoise of 0.36,0.39,0.42,0.45,0.48 for that last multidiffusion run, thereby 6 ever so slightly sharper, more detailed versions of that pics will be generated. The higher the value, the longer it takes and towards the upper end more and more errors creep into the image, like legs turning into torsos, additional characters or just faces randomly beeing generated onto background features and all kinds of other AI-shennanigans. But other parts of these higher denoise pics will end up even sharper, more detailed while still retaining the original subject. So in the next step I'll usually photoshop a compound image where I simply take snippets from lower denoised pics and add them in layers onto the best basic image I'll have gotten from the multidiffusion to cover up the AI weirndesses, then softbrush erase the edges of the added snippets to blend them in and if need be the next/final steps(s) will be inpainting the remaining bad details on that. The thing is I'm kinda aiming towards quality (not claiming to be successfull in that area, though xD) and at least somewhere around 2k. Because a) I don't just want to spam AI-schlock (and thereby contribute further to the resentment many folks have against AI - pics in general) and b) I've got some pretty big monitors, so its also just for my own enjoyment. xD ^^