character training unfortunately no longer works?

[pip-bln]

I started a small project in which I wanted to train 20 faces of members to create funny pictures for a private catalog.
So far it worked quite well for my first 7 attempts and with a few tries I was able to generate a suitable image for each face.
Unfortunately, this no longer works at all. The result is only pixel mud and the faces are not recognized at all.
So far my source material was 40 images in 512x52 pixel format, with 100 training steps as recommended. Now it says that the images should be trained 200 times each, which obviously does not work. The specification with the text encoder has always seemed quite random to me so far and unfortunately not very accurate usable. Therefore I had set it so far simply on 50 wich worked quite well. Now it works neither at 50 nor at 100.
Is there possibly a chance to get it working again so that faces can be recognized and trained?

[TheLastBen]

If you use the same settings and instance images as before, I don't see why you wouldn't get the same results

[DivinoAG]

What I don't understand is that before those changes, the lower text enconder value (40%) and 100 steps per image worked fantastic for faces. But now it doesn't work anymore. For example, this is an image created with a model I created on the 5th of this month using these settings I just mentioned. Features great resemblance to the subject, really good style transfer, it's all around a great model.

Last night I tried to make another model using almost the same 30 images (I replaced a couple), same number of steps per image, same text encoder value, and the person barely comes across the image when using the exact same prompt, the style is completely overwhelming the final image.

It's not just the change in the recommendation, but something clearly is not working as it was before in the notebook itself, or at least we would be able to reproduce the exact procedure we used... last week. It would be helpful to understand what changed, so we can adjust our settings accordingly.

[TheLastBen]

from the 5th to today, I haven't changed anything related to the training, you can check the commit history. It is just your settings, don't the use "contain_faces" option

[DivinoAG]

Why not? Isn't that the purpose of that option?

[TheLastBen]

that option prevents the trained subject from leaking to other subjects, but it doesn't improve the result and it requires way more steps because it has 200 more pictures to train on.

[pip-bln]

Honestly, it looks similar to DivinoAG for me - So even if nothing has actually changed, it has definitely changed in the result. I had as mentioned with my first 7 faces absolutely no problems and now I suddenly do not get further. My impression is that faces are no longer recognized as well as before, or hardly at all, but instead more the style of the entire image.