Well the training part of this lora went really well; the inapint part, not so well. This posture contains crossing limbs of 3 characters, is the most difficult thing I had worked with SD so far.

Strongly suggested to use txt2img only, very difficult to select the characters you want, results are best with horizontal, 1:1 also works, but need some luck with vertical canvas. By default, it can generate 2 boys 1 girl; if you have lora for boy and girls, it will most probably generate images with the lora according to the gender. You can also prompt to 3 girls, haven't tried 3 boys yet, not sure what it will turn out. Do notice this lora will generate big breast, otherwise you don't need the servants to hold them.

If you really want to control the character's placement like what I posted here, you will need a well generated img with the composition you want, throw that in as canny, open pose or depth, use latent couple in img2img, then tiled diffusion to enlarge it in img2img, slowly inpaint the defects 1 by 1. (Took me 8 hours to trial and error :/)

Finally, as always, go break some eggs, have fun!