Flux-European Beauty

V2

Trained on 600 handpicked and cropped photos for 60 epochs.
The training set doesn't include piercings, tattoos, necklaces and other jewelry.

Is V2 better than V1? To be honest, not really. The problems with text and teeth have improved a bit, mostly a bit sharper and more detailed backgrounds, but I'm not really satisfied. It still happens too often that the teeth don't look good and I just don't understand what the problem is. I have gone through the data set several times, it is absolutely high quality. And it shouldn't be due to the training itself. I've trained roughly 40 Lora in the last 2 months and I've never had these problems with any Lora, even when I've only trained on 512px.

I will upload some comparison images between V1 and V2. Then you can decide for yourself which version you like more. The differences are usually not very big anyway, although I have trained from scratch with 120 new images.

_______________________________________________________________________________________________

V1

Trained on 480 handpicked and cropped photos for 27 epochs.
The training set doesn't include piercings, tattoos, necklaces and other jewelry.

The actual goal was to train women from different European countries in a Lora in order to then prompt, for example, “woman from Ireland”, “woman from Germany”, and so on.

However, the idea doesn't work as intended and it makes almost no difference which country the woman comes from.

But the Lora itself has a good variety of different types of women, good quality

and in general I am very satisfied with this Lora.

I would be very grateful if you could leave me your feedback here
and post images with your results.

I would really appreciate it if you guys would leave me

a small Buzz tip here so I can train more Loras.

_______________________________________________________________________________________________

Recommended Settings:

• Distilled CFG Scale:

  • 2-3

• Sampling method and Schedule type:

  • I personally use DEIS/Beta.

    You have to decide for yourself if other combinations deliver better results.

• Steps:

  • 30 + HiRes

• Resolution/Aspect Ratio:

  • The Lora is trained on 1:1 and 2:3 aspect ratio.
    You will have to test if other aspect ratios work.

• Lora Weight:

  • 0.6-1

• Trigger Word:

  • No Trigger

• Prompting:

  • Does not require any special prompting.

  • Should work with any type of prompting.

_______________________________________________________________________________________________

Known Issues:

• Text:

  • Above 0.8 problems to generate text.

• Teeth:

  • I don't know why yet, but this Lora has problems with teeth.

  • I'm still trying to find out why and try to avoid it for the next version.