Thanks! Basically, like I mentioned in the description, i make a bunch of Flux gens and pick one I like. Then I'll run it through a bunch of img2img workflows, usually using SDXL to get rid of that classic Flux plastic skin until I'm at a point where I'm satisfied, then run the image through Kontext to create a dataset. Sometimes the dataset images need some img2img love as well, since kontext is flux so, even with loras, that plastic skin can start to creep back in.

@estiva_st thanks for the kind words! I use Comfyui's Flux trainer as well, but also I'll use civitai's trainer as well, for when I don't have 8 hours to devote my laptop to training. I usually aim for around 1800 steps; the only civitai default settings i'll change is knock the learning rated down to .0004 and constant scheduler, as that's what I use in comfyui.
The dataset is the only thing I can think of that really makes the difference. I typically aim for about 100 images if I can, but 50 is usually the absolute minimum. And I typically don't really care what the aspect ratio is of the training images. If a perfect square works, great, but a random size like 903x1221 wouldn't be out of place either. Most images will be just the face/head, but I always try to add in a decent amount of torso and full body shots if possible.
I hope this helps in some way!