[ 3 / biz / cgl / ck / diy / fa / ic / jp / lit / sci / vr / vt ] [ index / top / reports ] [ become a patron ] [ status ]
2023-11: Warosu is now out of extended maintenance.

/vt/ - Virtual Youtubers

Search:


View post   

>> No.64851930 [View]
File: 703 KB, 1664x1408, 00063.jpg [View same] [iqdb] [saucenao] [google]
64851930

will she be there

>> No.47694052 [View]
File: 703 KB, 1664x1408, DB5B49B6-91EE-49D0-A368-FDF8C48FF80D.jpg [View same] [iqdb] [saucenao] [google]
47694052

I need miyunny so bad anons you don’t understand. I’m STARVING.

>> No.47680112 [View]
File: 703 KB, 1664x1408, 1681086187545132.jpg [View same] [iqdb] [saucenao] [google]
47680112

>> No.47003360 [View]
File: 703 KB, 1664x1408, 2.jpg [View same] [iqdb] [saucenao] [google]
47003360

>>46977761
sure sorry for late reply
first here's a catbox of the final pic https://files.catbox.moe/jdbaqp.jpg

second the process.. i wanted to remove the background and upscale the pic so i started with a screenshot ofc, and chopped the background out in photoshop and then ran it through an upscaler (r-esrgan 4x+ anime6b) to increase the size so i had more to work with, and then fixed up some of the upscaling artifacts and painted out her shorts as well as i could and ended up with picrel

i would have been more than happy to just inpaint out her ass area but the linework got fucked by removing the background and upscaling and i wanted to hide the edits better so i decided img2img was the way to go. i fucked with denoising values and controlnet params that would be enough to fix my editing job but still preserve the most detail of the original pic

then i randomly happened to stop an attempt only a few steps in and noticed the result managed to fix the linework while not also slowly denoising away too many of the other details and fucking up the coloring. idk this is probably a known thing but going forward i'm going to experiment with modifying steps instead of just really low denoising for when i only want a little edit-hiding help.. usually i always leave steps at like 24ish for the dpm++ samplers and 40ish for DDIM while trying for very low denoising values

anyway the fine details still got fucked with a little but i was too lazy to fix them all. i copied the original hand/lollipop and her eyes from the original upscale over the final output and called it a day, since her fingers properly holding the lolli stick and the distinctive eye detail got mangled even on low step/denoise

last, as for the model and prompt i'm still on aom2 because i'm too lazy to try out the 50 different blends vtai has churned out to try to see which is an improvement and the prompt is mostly just a regular placebo preamble with a few relevant details about the actual pic slapped on the end. i've removed most of that quality bullshit and it doesn't make gens look appreciably worse but w/e

sry for long reply but i find this stuff fun

>>46997013
was considering my own scuffed lora training lol. if that's pure txt2img that's a good start desu

Navigation
View posts[+24][+48][+96]