Default appearance:
p4rise, kujikawa, rise, twintails, long hair, brown hair, pink earrings
glasses, pink-framed eyewearWinter uniform:
p4rise-winuni, yasogami school uniform, serafuku, long sleeves, black shirt, white turtleneck, school uniform, pleated skirt, houndstooth, thighhighs, zettai ryouiki, Summer uniform:
p4rise-sumuni, yasogami school uniform, serafuku, short sleeves, white shirt, skirt, pleated skirt, houndstooth, white sock, loafers,Summer casual:
p4rise-sumcas, heart necklace, pendant, dress, short dress, orange dress, halterneck, spaghetti straps, flower print, pants, capri pants, white pants, sandals, Summer casual 2:
p4rise-sumcas2, alternate hairstyle, hair down, hat, casual, t-shirt, black t-shirt, tank top, pink tank top, layered shirt, shorts, blue shorts, denim shorts, Winter casual:
p4rise-wincas, hoodie, purple hoodie, hood, tank top, pink tank top, layered clothes, skirt, pantyhose, Tofu:
p4rise-tofu, head scarf, kappougi, turtleneck, pink turtleneck, skirt, blue skirtSki:
p4rise-ski, jacket, pink jacket, long sleeves, white sleeves, hood, gloves, purple gloves, pants, blue pants, heart \(symbol\),Swim:
p4rise-swim, swimsuit, string bikini, orange bikini, flower print bikini, flower print, side-tie bikini bottom, halterneck,Swim (Golden):
p4rise-swim2, halter bandeau bikini, bikini skirt, pink bikini, polka dot bikini, frilled bikini, eyewear on head, sunglasses, navel, Timeskip:
p4rise-timeskip, alternate hairstyle, hair down, eyewear on head, sunglasses, bead necklace, layered clothing, off-shoulder shirt, bare shoulders, see-through shirt, see-through, long sleeves, tank top, rainbow tank top, brown pants, capri pants, bracelet, wristband,Rise Kujikawa, the navigator & Lovers Social Link in Persona 4.
As always I curated the dataset for this Lora from official artwork, screencaps I took from the P4 animes and portraits & screencaps from the game.
Her winter casual outfit is a bit inconsistent, but partially since it's just kind of a weird outfit? I think she's wearing a tank top over a purple hoodie dress thing? Anyway give it a go and let me know your thoughts.
Description
Experimented with training settings to lower filesize
FAQ
Comments (4)
Hi there! Since you mentioned the filesize, I'll chime in with a few tips (since I love the P4 characters dearly and that's what lead me to stumble upon this.
1) The main reason many Pony LoRAs are 200+ MB is simple: CivitAI's trainer sets the dimension and alpha settings WAY too high. These are buried in the advanced settings as "Network Dim" and "Network Alpha." I'm not sure what the defaults are on CivitAI (I have a 4080 Super, I don't need to generate or train on the site), but the general wisdom from folks "in the know" is that for most things that aren't overly complicated detail-wise, you can lower this a lot and still retain very accurate quality. The Alpha should always be half of the Dim, so something like Dim 8, Alpha 4 or Dim 16, Alpha 8 could work pretty well. If you feel like the model isn't quite right even after a fair few epochs, that would be when you'd want to raise them both to the next step and see if that would solve it. The old SD1.5 maxim of Dim 64, Alpha 32 is overkill for XL-based models, and it's what produces these 200+ MB LoRAs when that much information is simply not necessary.
2) In addition to not cranking the Dim/Alpha too high, there is also the option of resizing the LoRA, which you can do with Kohya's tools or with other (faster) tools. This essentially works in multiple ways (depending on what you tell it to optimize for), but the basic gist is that they will chop out data that strays too far from the model's baseline, attempting to leave only the good data that "matters" behind.. This means it's a lossy, one-way reduction in filesize - think of converting a WAV ripped off a CD to an MP3. If done carefully though, what you chop out isn't noticed - but the filesize is a heck of a lot smaller. I've seen sub-20 MB LoRAs that still resemble the character or artstyle pretty damn good. I've seen a creator release both this reduced model as well as keeping the original untrimmed LoRA up for those who want absolute quality - that might be a good middle ground.
Hope this helps! And I hope to see a few more of the girls in due time. I know there's a few in the game that were still missing from here last I looked (Ayane and Mrs. Kashiwagi come to mind).
Thanks for taking the time to type all this out! Yes I've begun to realize that the Network Dim/Alpha I've been using is way way too high. I experimented with I think an Dim/Alpha of 8/4 recently and I was shocked at how good the quality was compared to the ridiculously oversized versions.
I've also seen the resizing tools but have been a bit scared off since it's "destructive" as you said. But for a noncomplicated character lora it's probably not going to be a big deal, might as well give it a try!
A fellow Ayane enjoyer! Haha, Toriumi is my favorite teacher, Kashiwagi comes on a bit too strong for my taste :P But I'll see what I can do, it's a bit harder with the side characters since there's way less official art of them, and (ironically I know) I try to keep my datasets free of unofficial/AI stuff to keep the quality high.
@OneMoreLurker Thanks for the tip! :)
I'd say that really the LoRA resizing is mostly a case of "for those who are absolutely crunched for space." IIRC at 8/4 you wind up with about a 50 MB LoRA, and obviously, you can stuff the space of four of those into one more "typical" 200+ MB LoRA. As I said, a good middle ground could be to offer both - the untrimmed 50 MB LoRA or whatever, and then a trimmed-down one. But of course, it's your model, your choice - if you don't feel like going through all that, you certainly don't have to. :P
Oh, one other area where a higher Dim/Alpha might be useful: If you're attempting a multi-character LoRA, or possibly a multi-outfit LoRA (something P4 lends itself to quite well). If you ever do one of those, try and see if you can keep it smaller (smaller definitely makes it easier for the impulse downloads, especially if it's of good quality!), but if stuff bleeds too much and/or it's just not capturing enough details, a combination of better tagging, and if need be, better Dim/Alpha can help solve that. (Usually it's more the tags' fault than the LoRA size though.)
As for P4... ironically, my first playthrough of the game, I went with Yumi over her in terms of clubs, lol. But I do like Ayane as a character. The biggest knock that's against her is that they've already got a character who already fits into the kid sister role, and so once you strip away Ayane's self-esteem issues, there's really not much left to her thaat Nanako doesn't already exploit to hell and back.
Still, won't complain once someone gets a LoRA up. And yeah, I feel you on the "official art" deal. The games will definitely be limited on that, but she was in the anime, so that might help. Heck, I know a LOT of creators just make things direct from the animes, and as someone who always found official style-feeling stuff that much more impressive, this is perfectly fine to me! :)
@Cult_of_the_Black_Buffalo
It's definitely fun playing around with and squeezing as much performance into the smallest package possible. P5 Makoto was a good experience for me in doing multiple outfits and it's something I'd like to do more, but man curating the dataset manually takes a ton of time, especially since I'm trying to avoid doing booru scraping.
I'd like to do another P4 playthrough but after P3R this year the combat especially feels dated. Hopefully it's next in line for a remaster.
Also, Merry Christmas buddy: https://civitai.com/models/1070017?modelVersionId=1201052
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.