Isekai wa Smartphone to Tomo ni Girl Pack
Samples See Gallery below + Grid's.
v2.0 has all 1.0 include
Enjoy :)
Added S2 (v2) Girls:
Bell
Caroline
Fredmonica
Goddessoflove
Hildegard
Lucia
Prim
Reliel
Rosetta
Sakura
Sushie
Version 1 Girls:
Ael
Alma
Cecile
Charlotte
Ellen
Elze
Francesca
Lapis
Leen
Linze
Micah
Olga
Sushie
Tsubaki
Yae
Yuel
Yumina
Trained on: W:480 H:720
Triggerwords:
"name" - see list
anime art style - Trigger for Anime Style (Beta)
See Preview Images for my Prompts.
Preview Images generate:
LoRA Load over: additional-Networks (work without it)
Model: abyssorangemix2_Hard
VAE: Anything-V3.0.vae.pt
Txt2Img and Upscale settings:
DPM++ SDE Karras
Steps: 20
480x720
CFG: 7
LoRA Strengh: 0.7
Upscale: 2x
Denoise: 0.6
Upscaler: Latent
Description
FAQ
Comments (12)
I just pass by to say this is definitely what we should do.
Not sure if you have trained the model to make prompting multiple characters together possible though.
Each character has its own training with different steps, but it also takes much longer ~4-6hrs (RTX3070), also depends on the number of people, but you can combine them via prompts but its hard ^^".
Thanks for the hint i train in v2 with multiple girl in my dataset :)
@sysdeep My models can easily put two characters together but more are generally quite difficult. Training with scenes that contain multiple people definitely help for this and it is also possible to put characters that never appear together in a same image after training in this way (even with characters from two different anime series models after applying an add difference merge).
I am personally on 4090 so I can imagine training on 3070 can be much longer.
How do you go about training multiple characters under 1 pack?
Its in work for all my Anime Pack come a Group Pack DLC and nsfw DLC :D
Just leaving this here for anybody, a good general use weight for this is 0.8. This gives it some flexibility for if you want to change outfits or something, while maintaining the anime style. If combining with a second activity LoRa, or trying ot make them perform certain actions (won't judge), 0.5 works wonders.
Just leaving this here as well for some extra tips. some characters have weird hair ornaments on them like elze. when using 0.4-0.5 strength they can often be lost or come out wrong. idk how to get them in txt2img (I tried tagging 'hair ornament', it doesn't work). A work around is to get the low strength image and put it into inpaint with the same exact prompt but change the lora weight to around 0.8. use latent noise with a strength of around 0.8. draw around the point those weird bang holders (I think?) are and it should come out. same with every other character.
also if you are not doing nsfw scenes, only changing characters outfits and poses, 0.6-0.8 are much more accurate. 0.6 being less anime like, 0.8 being more anime stylized.
Really happy with this lora. in general lora multipacks aren't good for more dynamic stuff, and this is no exception. But it is giving me the idea of just generating 100 images with this overnight with a strength of around 0.8-1 and then making my own lora using this for a specific character and fine-tuned how I like it. Never made one before, but a man's gotta start somewhere.
@hentaireviewer69748 Thanks for your tips / review, this one of my first LoRA :D v2 will come with Season2 :3 this will work better :D
is there a decent way to dial back the art style while still keeping character features intact? for me it seems to be doing the opposite at times.
Your LORAs are great, but I am having a lot of trouble getting characters to appear on model with this one. It works well enough, but I really hope you have plans for a V2 that works as well as the other group LORAs you've done.
check v2 :3
@SysDeep The V2 works amazingly better. Thank you.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.