@JujoHotaru Thanks, I read the three pages and I know the principle of Copy machine Lora.

But I don't know how to generate a Lora like you about Face/Eyes/Expresion/.

I can't get any information of you lora such as Prompt and Training Set……

Can you share some training experience?

@JujoHotaru Surprising results, a single facial expression image material can achieve such amazing effects through Copy machine LORA?

@JujoHotaru Therefore, training the expression lora is to train the training set with expressions through Copy Machine Lora without writing prompt words.

Can we get the ideal Lora in this way? Are there any other steps?

@JujoHotaru I tested this very special expression @_@.

If only one piece of material is used, this image can indeed be generated very well due to oversaturation.

But when I lower the weight, I can't reproduce the expression very well. However, when I increase the weight, the prompt words do not control the other images.

How to solve this Problem?

@JujoHotaru I tested it and still couldn't get the desired results. For this Lora, can I see your training set and training parameters? What is the approximate LOSS value at the end of training?

@JujoHotaru If I understand correctly, it should be handled like this:

1. Train the original material to create an oversaturated Lora.

2. Merge oversaturated Lora and large models.

3. Make certain modifications to the original material and train on the merged large model until it is oversaturated.

4. Subtract two Lora models?

Finally, we get a experience lora, is this correct?

@JujoHotaru Currently I use one picture for training.

First, I removed the eyes from the image and trained a supersaturated Lora.

Then I merged it with the larger model.

Finally, I used the merged large model to train the original image and obtained a supersaturated Lora.

This has some effects, but the restoration effect is not ideal.

Maybe there is something wrong? I'd love to have a look at your training set and learn from you!

@JujoHotaru This week, I tested it many times and felt the amazing effect of this training method.

For example, changing the grayscale and saturation of an image is ideal.

However, I don’t know why, but some goals cannot be achieved well.

For example, I want to add a star to the eye, but I cannot get the desired effect through CopyMachine. The star is either blurry, or the star cannot be drawn at all.

Here, I really want to learn how you process the material set. Can you send a few Lora training materials to my email? [email protected]

Thank you so much!