这回训练的目的是，稳定生成两个角色，并且通过位置固定特征（画面左边的年龄更小）。
基础功能：可稳定生成两个女性
初级用法：稳定生成两个女性，表现出她们的年龄差
权重0.8-1，使用Hires. Fix，Denoising 0.3-0.4
进阶用法：使用两个角色lora，稳定她们的年龄，妹妹在左边，姐姐在右边（并不是100%稳定）
1.使用Hires. Fix，Denoising 0.3-0.4
2.同时打开latent couple和 Composable Lora。latent couple使用默认的矩阵即可。
3.使用lora插件（注意：必须是插件）加载本lora，建议权重0.3-0.5之间
4.按照以下格式填写prompt
“age difference, 2girls, girl on the left,
AND 1girl,age difference, 2girls, girl on the left,small breast,<妹妹lora1:0.5>,[lora1‘s trigger words]
AND 1girl,age difference, 2girls, girl on the left,big breast,<姐姐lora2:0.5>,[lora2‘s trigger words]”
要使用的两个角色lora需要适当调低权重。
需要找到本lora生效和过拟合之间的平衡点，建议使用xyz寻找。因为角色年龄受到角色lora影响，有时候需要配合其他prompt加大力度。
The purpose of this training is to generate two characters consistently and show their age difference through fixed position features (the younger character on the left side of the image).
Basic function: Generate two females consistently.
Intermediate usage: Generate two females consistently and show their age difference.
Weight range: 0.8-1 Use Hires. Fix and Denoising 0.3-0.4.
Advanced usage: Use two Lora characters and generate them consistently with the younger sister on the left and the older sister on the right (not 100% consistent).
1.Use Hires. Fix and Denoising 0.3-0.4.
2.Enable latent couple and Composable Lora. Use the default matrix for latent couple.
3.Load the Lora plugin (must be a plugin) for this Lora. Suggested weight range: 0.3-0.5.
4.Use the following format for prompt:
"age difference, 2 girls, girl on the left,
AND 1 girl, age difference, 2 girls, girl on the left, small breast, <younger sister Lora 1:0.5>, [Lora 1's trigger words]
AND 1 girl, age difference, 2 girls, girl on the left, big breast, <older sister Lora 2:0.5>, [Lora 2's trigger words]"
Adjust the weights for the two Lora characters appropriately.
Find the balance point between the effectiveness and overfitting of this Lora by using XYZ. Because the age of the characters is influenced by the Lora characters, sometimes other prompts need to be combined to increase the effectiveness.
Description
FAQ
Comments (6)
I always like your lora, but they tend to look very overfitted.
What model are you training on? try animefull-pruned model for training and aim for a weight of 0.8 - 1, use a karras++ sampler at 7cfg to review your images, not euler.
Thank you for your support. I usually use the original model for training. Regarding overfitting, using lora does indeed present this problem. Especially since I want to implement multiple functions within a single lora, this problem becomes even more severe. Today, after training, I spent six hours selecting this version of the model from thousands of images and found a relatively stable method for implementing 2 character lora. It took me a lot of time. There is no special reason for choosing Euler, it's just because it's fast. I will continue to optimize this model and try to reduce the problem of overfitting.Perhaps training with a different model is a good method, and I will try it next time. Thank you for your suggestion!
@romiolast euler is an ancestral sampler, it is not good to use them since they introduce a lot of variation and don't converge fast. Karras++ SDE is usually best, use with low steps 10-15 since it converges a lot faster and gives you a nicely saturated style.
if you use anime-full-pruned model for training it will be very flexible, if you use euler the image will be too weak, use Karras++ to get a nice image with good colors. 👍
@cyan8 Thank you. I have a lot to learn.
Nice to meet you, I am karouji.
If you can,could you upload your excellent models(particularly this) to TensorArt?
I want to use your models at there.
Sorry for asking, but what is the different between this LoRA and the other LoRA? (models are generally gender-agnostic) Both are two people with different height hugging or doing things so IDK about the details. https://civitai.com/models/21618/two-person-lora-lora-update-for-character-lora
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.