⚠️ ALPHA RELEASE WARNING: This LoRA is experimental. Expect quirks, and “rough edges” while I iterate on the dataset and training settings.
What it does:
Captures CivChan’s facial identity and overall look
Works for portraits and full-body prompts
Handles wardrobe changes
Known issues / limitations (alpha)
Not fully tested (you are the testers)
Recommended settings
Model: FLUX.2 Klein 4B
LoRA strength: start at 0.6–0.8, go up to 1.0 if you want stronger likeness (its probably underbaked)
Steps: 50 (base Klein often needs more steps)
Guidance/CFG: around 4–6
Resolution: 1024x1024 recommended; higher resolutions work 🤷
Description
Here are the tags we put besides the trigger, grouped so it’s easy to sanity-check:
Core identity cues
long pink hair(most common)shoulder-length pink hairpink hairEye colors:
purple eyes,blue eyes,green eyes,brown eyes,gray eyes/grey eyes
Outfit / clothing
maid_outfitshirt,topjacket,hoodie,sweater,turtleneck,tracksuitjeans,skirt,dress,blousesuit,tiearmor
Framing / camera-ish
portraitclose-uplooking at viewer
Expression / pose
Expressions:
neutral expression,serious expression,slight smile,smile,angry expression,surprised expression,sleepy,mischievousPoses/gestures:
arms crossed,waving,peace sign
Generic subject tag
woman
Comments (8)
Work also with Klein-9B...great Lora :)
That surprises me. That is not supposed to be possible.
It works very well. Tested with cfg 1-1,5 and 4-8 steps.
You can check the results on my profile. The last 2 posts i made.
I don't know why i don't see them here on your Lora page...
@Herastura Damn, those came out really well.
@ccollins thx :D
@Herastura if you could try the same prompt(s) with and without the LoRA active and seed locked that should tell us if it's the LoRA or the models innate knowledge. I think it is working based on the image likenesses, but I'd like to confirm that. Flux.2 does have some innate knowledge of CivChan.
@ccollins Just tested it (without Lora, fix seed, same prompt) and it's the models innate knowledge. The pics are to 95% the same...
I will make a Flux.2 Klein 9B version. It's much faster to train on 4B so Im starting here.