← sources

I trained a Krea 2 LoRA on 356 real portraits with a multi-parameter face vocabulary - you can prompt face features directly

source: reddit (reddit:StableDiffusion)published 2026-09-26 10:06ingested 2026-09-27 02:38

retained text

I use Krea2 for fashion photoshoots and editorials, but that same polished, inoffensively attractive face has been bothering me for a while. I wanted more realistic images with more control over the result.

So I trained a LoRA on 356 snapshot portraits of real people, and captioned every single one with a structured vocabulary:

* Face type: feminine, masculine, androgyne, baby face, tired, aggressive, gaunt
* Eyes: downturned, upturned, hooded, wide-set, close-set, deep-set, protruding, monolid, round (heterochromia works too - just name both eye colors and add some weight to the phrase)
* Nose: big, long, straight, tiny, short, bumpy, fleshy, roman, hawk
* Lips: big, wide, thin, narrow, heavy upper/lower, sharp, bow-shaped, down-turned
* Cheeks: full, hollow, large/low/high cheekbones
* Jaw: wide, narrow, soft, sharp
* Chin: pointy, wide, protruding, receding

At strength 1.0, with no keywords at all, the standard Krea face disappears and you get real human variation. Add the feature tags and you get exactly what you asked for - a bumpy nose stays bumpy, a receding chin stays receding. The more you describe, the better it works: tags stack well in one prompt, e.g. long fleshy nose, hooded wide-set eyes, narrow down-turned lips. This also works with full-body shots, but too many tags can cause some glitches

How I use it: strength \~1.0-1.3. I recently tested it at 1.3-1.5 and can’t go back to 1.0 now.

Details:

* dataset: 356 real people, European/Caucasian, age 20-40, 174 men / 182 women
* “full cheeks” + “round face” can give interesting body-positive results
* “tired, aggressive, gaunt” - these tags can produce some creepy images, be careful
* eyebrows! Please avoid describing anything related to eyebrows - they either appear too thick or don’t appear at all

Would be curious how it behaves in stacks with your favorite realism LoRAs - tell me what you try.

I’m also planning to improve this LoRA and fix the eyebrow bug, so any feedback is valuable

**Model:** [**Realistic Custom Face**](https://civitai.com/models/2911560/realistic-custom-face) **on Civitai**

document fetched 2026-09-27 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1wqmsmn/i_trained_a_krea_2_lora_on_356_real_portraits