Favorite character from my favorite show. More Arcane LoRAs to come!
Use model weight <lora:jinx:1> if you want images that could be stills from the show. Use <lora:jinx:0.65> for more flexibility in attire or style.
Check the image details for prompts and generation settings.
Description
Initial version
FAQ
Comments (6)
You need more pictures with just her face and cover up the x on her chest in the data set (Just draw over it). I made a lora of her in sd1.5 and it was more dynamic the way I did it, it also captured her likeness extremely well.
And train it on the word "woman".
Hey thanks for the advice. This is my first lora so I'm definitely looking for tips. I also noticed the 'x' was coming up a lot, so I'll try drawing that out in some of the pics and mentioning it in the captions. What is the purpose of adding more face pics btw? Does it improve accuracy or flexibility? Also last question, what do you mean by "train it on the word woman"? Does that mean just add "woman" to all the captions?
@Sextus_Rex120 Adding more face pictures means that her face will become a priority in most images. You have to remember, you are training a machine that will produce what it is trained on, so full body images are great, but for close-up shots you want the fine details of the face. For far away shots you want to take the images from her 360 turn around with the white background.
The more images there are, the more flexible the model, the less you need to train. I remember when I only trained on a few images, her side-profile didn't work out so well, but if you train on multiple images, zoom-ins of some images even, it will make sure to keep her facial details in tact and will also allow for different hairstyle prompting.
The image where she has her hands up and looks sad will improve her fingers and give her the appropriate nail colors, also the scenes where she holds the cup will improve her holding stuff.
Regularization images should be used to avoid "woman".
Have images with her mouth open, where she is angry and when she is sad and she will be more expressive.
The main point with having many images of a zoom in of her face is that you can change everything but her face to how you want it to be. So if you want morbidly obese Jinx it will be easier to trigger with many only face pictures, while keeping her facial attributes.
I also didn't use captions on mine, instead I used regularization images of women generated by the model and trained her on the concept "woman". For me, it triggered without needing to caption the images at all, it already knows what is in the image. I could write "blue twin braids" and it knew to trigger them.
So if you want her face from all angles with more diversity, add more images to the dataset and keep her face the dominant aspect of her. Even deformed faces like her shouting should be included.
I am not sure about captions though, check if the images are better with them or without them. Also make sure the training images have the 1024x1024, 896x1152 etc. ratios so that they trigger better (Unless Kohya has it built in, in the latest update). You can try without formatting too.
@Sextus_Rex120 Also you could remove the choker around her neck as well, or it will appear often. I found the choker always looks good, but if you want it to not trigger often, you need to remove it from some training images as well.
@virtualfix6885 Thanks for all the tips, I'm testing out a new model now. It's much easier now to generate without the 'x' on her shirt. The choker still appears but it's easy enough to remove it with a negative prompt.
I hadn't considered trying this without captions. I'll keep playing with it and see how that impacts the quality.