screenshots taken from a shitty overcompressed low quality bootleg youtube video so don't be too surprised if its complete shit
funky facts: this character has had only 1 rule 34 image produced in history and that was in 2004. in the 21 years since, nobody has given a single fuckin shit about this character. but that's what ai is for!
Description
Comments (4)
The lines at the bottom of your images suggest you print screened the video and got the YT interface in them. There's browser extensions that will capture clean screenshots of YT videos and avoid this issue.
Otherwise, try tagging your training data "blurry, lowres" as appropriate. I've had it work for Ellie and Atsuko. May also be able to use a 19?0s (style) tag depending on when it's from.
Edit: IMDB has a cover in decent quality. An image search also finds what appears to be a VHS label that's also pretty decent quality. Cleaning her out of those and including them would help greatly, as it would help teach what is and isn't "blurry, low res".
Edit 2: If you retrain this, I'd recommend using a trigger word (instead of trying to overwrite 1girl) and use "print shirt" over "multicolored shirt"
for shizzle
@scottstupp Huh?
When I was looking for something else I found the animation studio behind the series has some high quality pics and a video of a scene from the series in somewhat better than potato quality on their official website. Figured I'd note it incase you ever wanted to make a V2.