For those who are tired of identical perfect Krea faces
This LoRa is learned on 356 real people snapshot portraits, each portrait was described by up to 29 parameters
At strength 1.0 without any keywords the standard Krea face disappears, the real human beauty starts to appear. Using a standard description of facial features you can customise the person's appearance in extremely small details such as jawline, nose shape, lip size.
The more you describe, the better is result.
Have fun with the prompting!
LoRa usage and prompting tips:
Faces total: Feminine, masculine, androgyne, baby face, tired, agressive, gaunt
Eyes: Downturned, upturned, hooded, wide-set, close-set, deep-set, protruding, monolid, wide, round
Nose: Big, long, straight, tiny, short, bumpy, fleshy, roman, hawk
Lips: big, wide, tnin, narrow, tiny, heavy lower lip, heavy upper lip, sharp, round, bow-shaped, down-turned
Cheeks: Full, hollow, large cheekbones, low cheekbones, high cheekbones
Jaw : wide, narrow, soft, sharp
Chin: pointy, wide, rotruding, receding
LoRa Strength: Depends on your prompt. Sometimes it works great at 1.2. Ugly faces may appear if you increase LoRa strength up to 1.4
Heterochromia: Decsribe two eye colors and add "he/she has heterochromia"
The dataset is mostly european/caucasian men (174) and women (182), age 20-40
Created for high fashion unconventional beauty editorials
Works best for creating a character, then use the output with other models
kept the misprint in LoRa name, easy to find it
Description
Original
FAQ
Comments (7)
It definitely looks like your lora moves the image away from "inoffensively attractive" towards more of a genuine "fashion model with interesting features" look. That's an achievement in itself. I wonder if, combined with gritty realism loras we can get some averagely "ugly" people too.
EDIT - Haha! As per usual, I forgot to say "Thanks" and "I'll enjoy trying this one out".
This LoRa is actually trained on mostly "ugly" faces, but Krea does its thing and smoothes that out to "interesting features" level
Thank you for your hard work! Could you please share the tags/keywords you used? This is crucial—without using your specific standardized terminology, we just can't get the desired results.
Thank you for your feedback!
Some examples are included in the sample pics, basically you can describe the face in identikit style. This LoRa needs testing, I'm testing it out as well, and ANY feedback is very welcome and valuable
For instance, eyebrows - I defined eyebrow thickness (Thick / Medium / Thin / Sparse) and form (Upward / Straight / Rounded / Steep arch / Arch) And for now i see that the thickness parameter is way too sensitive: in order to get average-looking eyebrow you must write Medium, otherwise you get extremely thick or no eyebrow at all; And the form parameter seems to be completely ignored
I'm thinking of making a custom node where you can switch things and adjust intensity
Hey! I've added terminology info + uploaded sample images with visible prompts
Does this LoRA change Asian faces into Western ones?Are the Asian facial features preserved?
Actually, no. There were a few asians in the dataset, less than 5%, so when you type "monolid eyes" you can get a good asian look. But I can see it makes skintone lighter in most cases.
Here's my recent Cyberpunk style experiment, i got lots of asian faces with no specific tags, maybe the 'cyberpunk' word was a trigger there https://civitai.com/images/143876341