I'm glad you like them. I use very basic prompts, most of the details come from Adetailer and ultimate SD upscaler. Prompt example's: "Close up portrait of a beautiful woman in a library." or "beautiful detailed photograph, wearing a dress, standing in bookstore, turned to the side, looking at the viewer." I rarely use more than a sentence or two in flux for portraits.

I use the base ai-toolkit sample config mostly with a couple small tweaks.

A lot of creators will use a photo base of 10 - 15 photos at around 2200-2400 steps, which will get you great results. Not to mention much smaller file sizes (18 to 36 mb) and much faster processing times. I take things a bit overboard and target 100 min photos. I like the extra flexibility the extra sample size provides. I typically target 4,000 steps, which again is a bit high. But you can monitor the samples in the ai-toolkit to see if your model is getting over baked and use a lower epoch model if needed.

I save every 200 steps so I can pivot if needed.

save:

dtype: float16 # precision to save

save_every: 200 # save every this many steps

max_step_saves_to_keep: 6 # how many intermittent saves to keep

I create sample batches to review every 600 steps.

sample:

sampler: "flowmatch" # must match train.noise_scheduler

sample_every: 600 # sample every this many steps

width: 1024

height: 1024