You're welcome and thanks for the positive feedback 🤗

Thank you so much for such a positive review, it's really nice to hear that😉

Yes I already answered that in one question in the commentariat:

It was trained on basic PonyXL, and the customizations were one of the old standard presets in Kohya for SDXL (I can't find them in the latest version). In short, it's adafactor Scheduler and Optimizer. Text Encoder and Unet learning rate 0. Network Rank (Dimension) 128\128 alpha According to my observations it is much easier to get even with fewer steps with dynamic learning rate than with cosine or linear. And the most important thing is the dataset. For my personal needs I usually have enough even 1 image in the dataset for 100 steps. The best way to style it is to make a crop on the bust of a character with minimum clues. And it is better that the original image was at least 1.5 times the resolution per one than 1024. So during training much better retains the necessary details.

If we talk about this particular loru had to make several versions after which I deduced which images are perfectly understood by the built-in SDXL decoder and already well convey the style. Usually I take 2 portraits in high resolution (more than x1024px) and a few images in close-up. The main thing is that the images should not have too much complex visual noise for training (which can be quite difficult to achieve). I usually do between 1 and 4 images. In this lore I stopped at 4 images of 100 steps each.

There was another experiment with a large number of images and tagging individual concepts in the lore on the style of the artist bonifasco, which gave also excellent results, but had to painstakingly tag each of the almost 60 images.

I will also warn that the output model at a network resolution of 128 will weigh almost 1gigabyte. So now I always compress them through kohya with the help of non-dynamic method to a resolution of 8, which eventually gives an output 50mb file at 99% similar in quality to the original file.