@MiraBelle666 Dont worry, it's time for me to reuse my pony merge 😋

@MiraBelle666 just checked your training data. The amount of images is fine, but many of them come in sets of very similar images. You should probably use something more varied, same goes for poses.

Try to add more shots of her from different views - from side, from behind, worm's eye view(aka from bellow).

Try using different range of emotions and/or actions. All of this will help you make a Lora more flexible.

The amount of steps is probably rather low for such amount of images. Try to do something between 600-800 steps, maybe even 1000. You can also try to use more epochs, but this one is tricky. Personally, I find that for about 40-50 images 10 epochs and 800~ steps is ideal, but sometimes it doesn't work and Lora turns out rather rigid in terms of styles... still trying to find out a reason for it.

As for the baked in style, you can lessen the effect by adding images of varied styles in the dataset. With that I can gladly help 🤭

@MiraBelle666 you're welcome 🤗 but again, it may not work in baked in style department - I am currently testing different training settings on Loras that refused to cooperate. They're fine and very flexible with different prompts, but baked in style is still there. When/if I figure out what went wrong with them, I will let you know 😉

All in all, experiment! That's the best way to learn. Try to tweak the training to see how it responds, what is to your liking and what isn't.