>>24695 Haven't really tested it but it seems to have turned out pretty well https://rentry.org/cawt333g#tsukushi-akihito

>>24699 okay yeah that's already way better than the sd 1.5 versions you made. narehate countershading bleeds through sometimes but that's nothing that can't be patched up with the right negatives. spoiler for shotacon. guess i should probably re-train nanachi for pony, might work exceptionally well with the mixed tag space...

bros she's doing the handsign the korean feminists have infiltrated my local install its over

>>24700 That would be great if you do end up retraining Nanachi. I haven't had the best results so far with character loras on pony but I was thinking I'll probably try Faputa since I've got manual tagging for danbooru and e621. I'll just have to find a better way to combine them

>>24696 Do you still do a quality pass for the tags or did ai tagging get good enough for it not to matter?

>>24703 It's been good enough for styles because even if they're not completely accurate it doesn't really matter. Auto tags are better than none and it's better than pulling them from a booru in a most cases. Accurate tags do matter for character details or concepts and auto taggers aren't super consistent. They're a good starting point and again better than no tags but you will see a noticeable improvement with manually removing incorrect tags, making sure the same tags are consistent and present across your dataset, and adding more tags if you think they are important. The whole pruning inherent character details is also dumb but that's a separate issue.

>>24704 Do you also try to limit the number of tags? I remember back then you had to set up a token limit, and I don't really know what happens if your tags exceed that. Do they just get cut off? Will that degrade the quality? >The whole pruning inherent character details is also dumb but that's a separate issue. Huh really? I did that and was pretty satisfied with not having to tag obvious stuff like cat ears or eye/hair color. But then I only ever trained one lora. Also I autistically tagged every outfit detail and that didn't turn out very good as there would be conflicts with other outfits that shared the same tags and I would have to tag absolutely everything to get the outfit look somewhat close to what it's supposed to look like. Making up custom tags for the outfits also didn't do anything. Speaking of custom tags, do you still train with a custom trigger word for characters? With my lora it turned out to be pretty useless, and I think the gens looked better if I didn't use it. Maybe because every image in the dataset had 1girl, so that turned out to be my trigger word.

>>23690 haven't trained since around the middle of last year and noticed SDXL wasn't a meme anymore. been using kohya sd with a powershell script but it's probably outdated by now. is adam8bit still the way to go?

>>24707 yeah adamw8bit is still good for sdxl. some people use prodigy but i feel it tends to undertrain. train on base pony and make sure your scripts are fully updated. a few anons have shared training settings in this thread that are quite good, poke around the MEGA links and see what looks best to you.

Damn, I posted the wrong image.

>>24707 >>24708 trying to train a lora on pony using an updated powershell script. keep getting this runtime error, which doesn't occur when training on fluffyrock or other sdxl models. any ideas?

>>24711 nvm was retarded and forgot fluffyrock wasn't sdxl based

no XL loras of Alphonse(White Datura) or As109?

>>24713 There's an as109 lora for pony on civitai, in my experience it requires adetailer face to function

>>24713 I've been working on a Alphonse Lora but I haven't gotten to look how I like yet. It either doesn't look like his work when I go for lower steps or it fucks up the hands when I go higher.

>>24715 >or it fucks up the hands when I go higher Happens even on based68 if you use the built in tag without mixing it with another artist lol.

>Banned from Pixiv over near official looking Pokemon pictures. I flew too close to the sun. Didn't realize it wasn't obvious that... well, it's not official considering the content. No warning, nothing. Just instantly fucked over.

>NAI sub is finally done Probably for the best. I became addicted to that shit.

>>24718 It's over for me whenever v4 is released

vibe transfer actually works pretty well

God I wish that was me

>>24722 how does it work? most i can manage is fumbling around with extremely low weights so it doesnt screw up the actual image.

>>24724 it pulls color palette and art details strongly and poses less so but it can be easily overridden by artist styles. i'm currently thinking of it as a third leg - style/subject/vibe transfer, and you can use any two to bridge over the third. might be good for pulling specific periods of an artist's drawing rather than fucking around with year: or costumes that don't have a tag, or characters that pull an official art style too hard. oh shit i might finally be able to recreate that creepy photo-edited silent hill 3 heather porn style

>>24725 >might be good for pulling specific periods of an artist's drawing rather than fucking around with year That's exactly what I've been using it for. Also if there's a style that doesn't quite work or look right with emphasis, just use vibe transfer and set the reference strength to around 0.2-0.3

>>24725 er style/subject/theme, and you can use vibe transfer to paper over one that isn't really working anyway i just started playing with it again and it feels like there's potential here, think it's gonna involve a lot of twiddling quality tags >>24726 good to know

sadly not strong enough to carry a character that the model doesn't know

>>24718 Time to get addicted to pony/autismmix.

>>24729 Where do I start?

bread time

>>24730 I guess you can give pony a try first like I did, but I wasn't really able to gen anything I like with it. Well I also had to play around with autismmix for some time to get results that I actually like.

>>24732 Original cup poster. I tried using pretty much every pony mix, but I just can't get results I like. Maybe I need to fiddle around more.

>>24733 The style tag combinations here https://github.com/Siberpone/pd-styles/blob/main/styles-pd-v6.csv + style LoRAs is what made me happy with my autismmix gens.

New bread >>24735 >>24735