Updated for differen triggers/more epochs - twice as many total steps so the character is much better known.
However, both versions work well, and which gives best result varies with prompt.
Trigger: ChunLi
Outfit triggers have changed:
classic blue dress, pantyhose, pelvic curtain
sleeveless blue halter top, blue-full-leotard
spiked bracelet
Captioned fields of view: upper body, thigh-shot, full body.
Decided to attempt training two different outfits, to see if Flux could handle it on a low dimension (Network Dim 4).
Classic outfit: Chinese clothes, Spiked Bracelet, Pelvic Curtain, Pantyhose
Sleeveless version: Sleeveless, Tight Leggings
To do for V1
Recaption outfit parts
More Epochs
Increase Dim slightly
Description
Changed some things around to try and get easier prompting
FAQ
Comments (6)
Nice work on the two outfits. I had quite a bit of trouble with it. How did you go about tagging it?
I've been meaning to write up a short guide on my training experiences with Flux, including this - but for this, here's how I ended up doing it (of note, try to keep your dataset about equal between the outfits):
- Have a distinct token as a trigger [same as with SD/XL/Pony]
- You can start with WD14, but it's only to do things such as prepend your trigger, and help you pick your outfit description; I let it add a max of 5 captions, with a confidence of 0.5 at least, as my starting point.
- Prune/Remove any tags that have to do with physical appearance (hair color, eye color, length, the buns in Chun-Li's case, jewelry if its ubiquitous, etc)
- Pick the outfit elements you do want to be able to control and try once again to have the word sequence be specific: this is where I felt the need to retrain a second version, and it likely can still be improved. I kept it to 3 terms max + the Spiked Bracelet which was in nearly every picture. Merge tags together or remove extras until you're left with just those and the trigger.
- The rest of the tagging is optional; for this one I did do portrait/upper body/full body and 1-2 words on background, but that shouldn't affect multi-outfit. Tempted to try it with someone else to see if I can do 3 or more next!
@Alienor_ Nice that is similar to a lot of my own findings. My biggest issue was making the two outfits, did increasing the dim help with that?
@JaneB I tried both sizes, but kept 4 because I wanted to keep some of the detailing on the outfits (gold stitching). It was working with two, but I didn't end up posting it because it required more supporting tags to get the outfit "right"