This is a LoRA of an original character named Gemi. I aimed for a natural, approachable face like a woman you might see on the street,
Gemi is a fictional character created with AI and does not represent any real person.
GemiというオリジナルキャラクターのLoRAです。街で見かける普通の女性のような、自然で親しみやすい顔立ちを目指して作りました。
GemiはAIで生成した架空のキャラクターであり、実在の人物ではありません。
Description
FAQ
Comments (17)
This is really good lora. Thank you!!! and I hope you can make more :) / Can we know your traning tool?
Thank you! Glad you like it 😊 I used AI Toolkit for training.
@tigertakato Nice work. Was your dataset images and video? I want to try just images...
@fauxtographypro Thank you! The dataset was images only — I didn't use any video. Training with just images works well with AI Toolkit, so feel free to give it a try.
@tigertakato I've been trying all day with the help of claude and I can't even get it to train. I don't get it.
Ito
May you kee-p making loras!
@brnfd24434343d Thank you so much! I definitely will 🙏
Can you share your AI Toolkit config? Your lora is amazing!
Thanks! It's hard to share the full config in a comment, but the important part is simple: turn off Auto Frame Count and Do Audio, and change Num Frames from 39 to 1. That's it — everything else is the same as a normal LoRA training setup.
@tigertakato thanks for the advice. I'll try that out. I was having better luck training characters and styles with images surprisingly versus videos but still wasn't getting the results I wanted. I'll try your advice. That said, videos were necessary for training camera movement loras for the obvious reasons.
Could you provide some video samples you used for the training? Were they all crop video face shots or a mix of random body shots?
@ProvenFlawless Thanks! I actually trained on images rather than video clips. The dataset includes close-up face shots, of course, but also bust-up and waist-up shots. I also added some shots with more angled views (turned to the left and right) so the model doesn't get locked into a straight-on face.
@GlowingGuardianGirl 出来が良く感じなかったので一旦削除しましたが、再び載せます
@tigertakato Interesting. Question: Theoretically if I trained a lora based on a celeb (as you know, H3 has a problem with distance shots--faces become slop unless you increase the steps) could that correct H3 identity even further? As H3 already knows celebs and other people, would that therefore stabilize the face and increase face consistency? Say if I make a lora of tom cruise. Could that conflict or enhance the identity?
@ProvenFlawless それほど詳しいことは、私ではわからないですが、理論的には可能だと思います。ただ、有名人の顔の学習はNGですね
Never seen this kind of women in the streets in my country. Lol