I agree, out of all the vaguely "realistic" Illustrious models this is the only one that doesn't appear to have lost any of the original concepts from Illustrious.

@kftiger If you can fine tune models or merge them with loras that you've trained, I'd recommend doing so with a good set of images. Even 2-300 with studio quality lighting, and use 1536 res to train/fine tune, as opposed to the typical low 1024 mainstream standard. You may also get better results if you train in some landscape images, maybe a few hundred or so just to give the model a better understanding of non-square images.

And with relatively low numbers like that, you can tag them well(since it's realistic to go through and manually edit tags, as opposed to when training with 16k+ images). Another thing, training people, do men as well as women, please (if you do). Cuz males are severely underrepresented in the SDXL based models (including illustrious), to the extent that it doesn't even know the difference between dick and balls.

But, if you do take my advice and train people, Onetrainer has a useful tool for masking images. Mask the people using one of the people models from the dropdown in that tool, and remove any info about the backgrounds from the captions, also, make sure to use the 'masked traiing' option when training. Only caveat, you may need to go over the image masks manually after, cuz it often masks out stuff it shouldn't.

But, in my experience, the masking is better, cuz it creates a lora that knows what it's focus is, and isn't saturated with arbitrary background information. So it's focused and precise in contrast to non-masked.

Oh, one more thing, with people, mask out clothes as well. Cuz one thing I found loras have a nasty habit of doing if there's random clothings in the training, is to mash them all together during generation and create really ugly clothing/sheets/curtains, and it'll put those shabby cotton, and/or patchwork blends in even when not prompted in any way.