For the current version, I decided to completely rebuild the dataset, as before I took my generations, which I found successful, and as labeling files used directly prompts, which I laid down when generating (tags format). And it turned out to be a complete nonsense. But I didn't stop there and used the auto label function..... and the result was even worse. But I kept going! I passed each image through the “img-to-prompt” and this time I used the label in caption format. and this time I was satisfied with the result.... Although there is still room to grow.

What we have in the end, this LORA helps to create semi-realistic images in fantasy/sci-fi genre.