Weight: around 0.7-1.0 for v2face and v1; 0.4-0.7 for v2dress
Works fine with both anime and realistic (tried on deliberate2)
For other LoRAs better to use v2face version than v2dress. Theres some training image examples on the v2dress tab.
Main costume: pink hair, buns, blue jacket, green skirt
School uniform: pink hair, buns, white shirt, sweater vest, green skirt
Happy sugar life - The series is centered on high school girl Satō Matsuzaka (Satou Matsuzaka) who befriends a mysterious little girl named Shio, who was abandoned by her mother, and immediately becomes highly attached to her. They both agree to live together in Satō's apartment, and the former vows to protect that feeling of love, even if it means committing crimes or even killing people.
Description
Trained at 512x512, Anythingv3, 28/32 epoch, lack of manual tagging
Flip augmentation is on by accident
FAQ
Comments (16)
Before I even try, are any of her yandere expressions in there? It's more fun if you can make the unhinged pink haired girls actually be unhinged, after all.
There's 10+ images with yandere face in dataset, almost all of them are close ups. I tagged them with the glowing eyes tag. But only part of the generation is with a maniac's face, not a normal one or deformed one
I'll try to add full-body art this weekend or next weekend and assign yandere tag so better to wait till update
@sabin Tyvm!
@sabin where do I find this model to use with it? It doesn't work with abyssorange2_Hard.
ssdmodels_Anything-v4.5-pruned-mergedVae
@realitythe2ndecflgi690 i've downloaded it on huggingface
https://huggingface.co/Airic/Anything-V4.5
Also i've tried AoM3 and several other mixes.
May you post a link to your checkpoint that doesn't work? To the exact model and version
@sabin Here is the one that didn't have great results. https://civitai.com/models/4451/abyssorangemix2-hardcore
I have now been working with Anything V5 and it's been turning out pretty great. I will need to test it out with some other loras to see if I can get her in other outfits and situations for funzies.
@SkyLordAer i've downloaded aom2hard and dont see any issues https://civitai.com/posts/160552
@sabin Understandable.
I tried out the version with more outfits and it's working amazingly. Can you tell me more about how you got such good results? I'd love to be able to train my own loras soon, and most character loras I've tested are not this versatile. This one is honestly great.
@SkyLordAer This is my first attempt at training more than 2 sets of clothes. I wrote in captions "satou, a girl dressed in <costumename>, <costumename> costume, <othercaptions>". If only a small fragment was visible, such as part of a shirt collar i did not tag that costume. Something like brown jacket, green skirt also tagged. In general, I must try on a small dataset to test different options, how best to write the clothes tags. You can download training data examples https://civitai.com/api/download/training-data/50201
Also, I still can't figure out if specifying position tags in the frame improves the result or not affect it. These tags are
front view, side view, rear view (from behind), distant view (full body), close view (portrait)
//also i've seen tags like <costumename> outfit, now i think its better to use than <costumename> costume
@sabin Thank you! Can't wait to see how things go :)
@sabin To make it work, do you need to have different clothing training in separate folders within the same lora? If I wanted to do that with my own, could I split it up into folders and have it still work?
@SkyLordAer I didn't notice any difference whether there is a folder separation or not. Most likely the folder name is only important if there are no captions files. However, each folder can be assigned its own number of repetitions. It is also a little more convenient to work in dataset-tag-editor.
How exactly to divide the dataset into folders is a matter of taste. However, when using non constant scheduler like cosine_with_restarts the order of files is of some importance (high learning rate will be on different images) and it can be affected by renaming folders as well
Check this guide https://rentry.org/2chAI_LoRA_Dreambooth_guide_english
@sabin I'll check the guide because I didn't understand a word of that :')
Satou - it's very good, but also i want Shio :)
I want to try to make a model with several characters. Shio, an aunt, a blond guy and others.
I'll probably do it in 2 or 3 weeks. And I'll probably post the updated Satou model by the end of the week.
@sabin That would be amazing! I'd love to make some sort of comic about how they got to ave their happy ending. Just a soft comic about the living a peaceful life in the countryside. I know their "love" wasn't right, but some part of me just wants to see them happy man.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.