Original character LoRA for Krea 2. One woman, not a style pack and not a celebrity likeness.
Trigger: yarapxbr
Put the trigger first. Then describe the shot, the clothes, the place, and the light in plain English.
Who she is
Yara is an original character. Brazilian mulatta, mixed heritage, golden-tan skin, no tan lines. Round face, soft cheekbones, round chin, snub nose, large almond eyes, full lips. Extremely slim athletic swimmer build. Tiny perky chest. Narrow waist. One intricate black tribal tattoo on the left shoulder and left upper arm only. No other tattoos.
She is short. About 160 cm. That is an adult height in several regions of the world. Write her as a 24-year-old adult woman and keep saying adult. Thanks to the new identity workflow driven new dataset I could resolve the issues with v1 that she was rendered too young without mentioning age and "adult face" all the time.
Default hair is a short asymmetrical pixie: natural jet-black with vibrant purple streaks and highlights, sharply shaved left temple, slightly longer tousled strands on the right side. You can change the haircut and hairdress in our prompts, though.
She is an adult OC.
What it does
Holds the face across close-ups, three-quarter portraits, and full-body shots when the trigger is present.
Keeps the short swimmer build if you give the prompt room to say so. Medium and full-body frames need an explicit body line and a scale cue. A face-only prompt will not invent the 160 cm for you.
Keeps the left-shoulder tiki tattoo when that shoulder is in frame and you name it. Hoodies, coats, and high necks hide it. Do not mention the tattoo when the cloth covers it. Flip is off. If you flip the image the tattoo walks to the right arm.
Accepts wardrobe changes. Street clothes, scrubs, swimwear, gym kit, a wedding dress, and nothing at all all work. Describe the garment. Do not assume a default outfit.
Follows natural-language captions. Krea 2 does not want Danbooru piles. Write sentences.
Works on Krea 2 Turbo for fast previews and on Krea 2 Raw when you want the slower pass.
A second woman in the frame is just a prompt. This file only knows Yara. If you pair her for example with Zola, put
yarapxbron Yara only andzolacrmlon Zola only. The height gap is the point. Say it. Combining her with other OCs works better when using JSON prompting techniques.
What it does not do
It is not a universal beauty LoRA. Without
yarapxbryou should get ordinary Krea 2 output, not a soft Yara.It does not lock a single art style. The set is photoreal editorial. Illustration, anime, and heavy stylization are on you and on whatever other LoRA you stack. No promises there.
It will not keep two different faces in one frame as two trained identities. Only Yara is in the adapter.
It will not invent extra tattoos. If a chest piece, a neck band, or a right-arm sleeve shows up, you imported it from the prompt or from another LoRA. Name the single left-shoulder tribal and stop.
Backgrounds and lighting from the set can leak if you under-specify the scene. If you keep getting a park at golden hour or a warehouse lamp, say the room and the light out loud.
SFW and NSFW
It runs both. Portraits, street clothes, nude, and explicit scenes use the same file and the same trigger. No useful limit on heat. Prompt what you want, except one thing: Keep her an adult. The new dataset I used contained dozens of styles, clothes, moods, poses, and a couple of NSFW images including hardcore, so the model is as versatile as you can imagine.
How to run it
Trigger first, always.
LoRA strength: 0.70 to 0.85. Start at 0.7. Rise toward 0.85 if the tattoo or the pixie goes soft. Do not exceed 1.0.
Krea 2 Turbo: 8 steps. Guidance 0 to 1. LoRA 0.7. This is the daily driver.
Krea 2 Raw: 28 to 52 steps. CFG around 3.5. Same LoRA range. Use Raw when Turbo smears the mouth, the shaved temple, or the tribal.
Sampler / scheduler: Euler or Euler Ancestral with Simple on Turbo. Match whatever already works in your Krea 2 workflow. Do not invent a special scheduler for this file.
Resolution: everything Krea 2 supports. The dataset was based on 2k images completely.
Flip: off.
Prompt shape:
yarapxbr, photograph, 24-year-old adult, short dry tousled asymmetrical pixie cut in natural jet-black hair with vibrant purple streaks, shaved left temple, [shot size], [clothes], [place], [light], eye-level camera at her standing eye height, slight low angle, no high angleName the left-shoulder tribal only when that arm is visible. Swap dry tousled for wet messy when she is in water.
Krea 2 mostly ignores attention weights like (word:1.5). Say the thing. Do not wrap it.
Negatives do little on Turbo. Steer with the positive prompt.
Example prompts
Portrait:
yarapxbr, photograph, 24-year-old adult, short dry tousled asymmetrical pixie cut in natural jet-black hair with vibrant purple streaks, shaved left temple, tight close-up of face and neck, quiet unreadable expression, camera at her eye height, slight low angle, soft window light, pale wallStreet clothes:
yarapxbr, photograph, 24-year-old adult, short dry tousled asymmetrical pixie, shaved left temple, left-shoulder tribal tattoo, three-quarter figure mid-stride on a city sidewalk, black plunging jumpsuit, eye-level camera at her standing eye height, slight low angle, late afternoon sunWet hair:
yarapxbr, photograph, 24-year-old adult, short wet messy asymmetrical pixie, shaved left temple, left-shoulder tribal tattoo, mid-freestyle stroke in an outdoor pool lane, purple one-piece swimsuit, no hard nipples through the fabric, face turned for a breath, eye-level camera at the waterline, slight low angleWith Zola in one frame:
photograph, two adult women standing side by side, clear height difference, yarapxbr on the short 160cm woman with the black-and-purple asymmetrical pixie and the left-shoulder tribal, zolacrml on the extremely tall willowy woman with the honey-caramel baby-bang bob, both facing the camera, eye-level camera set to Yara's standing eye height so Zola reads tall rather than Yara reading small from above, city sidewalk, late sunPut each trigger on the woman it belongs to. Do not hang yarapxbr on Zola.
Training
Same recipe as the Zola and Blanka Krea 2 LoRAs. Different pictures.
Base: Krea 2 Raw
Type: LoRA, rank 32, alpha 32
Dataset: 53 captioned stills, one subject
Captions: natural language, trigger first, hair state written on every file
Images were built to a fixed identity sheet with a sophisticated workflow, then captioned by hand. Face, skin, tiny chest, and the single left-shoulder tribal bind to the trigger. Hair state, clothes, framing, light, tattoo visibility, petite scale, and anatomy were written out so they stay promptable.
Description
Complete overhaul with a new dataset
FAQ
Comments (7)
What did you change in the second version?
The woman represented by the model is now a tick older (27 instead of 21), which reduces the risk of underaged output to zero. I trained the model with a broad set of different clothes, styles, sceneries and emotional expressions, around 85% SFW images and 15% NSFW. So the model is as versatile as possible. I tested it in connection with many different other LoRAs (except other OCs, because of the still existing issue of OCs bleeding into each other) to make sure that her unique visual style stays intact even when combining her with fantasy/scifi themed stuff, or techniques like painterly, sketched, etc.
Two of the most important changes are that the face and the tattoo on her arm were 100% consistent across all the training images of the dataset. If you try both versions with the same prompt, you will instantly see the huge difference.
SenDai_Sempai, you can easily get around the age restrictions by lowering the LoRA weight to 0.45—though honestly, that’s a pretty pointless move, dude. Boosting the response to emotional tokens, on the other hand, is great. You might find HAELE 3D useful as a generator for emotional references; check it out—it could be worth a look. The only emotional LoRA that actually works somewhat is this one: https://civitai.red/models/2829908/detailed-emotions-and-expressions. It helps a little, but not much. That LoRA has a very narrow effective range; it stops working past a weight of 0.5 and instead starts heavily altering the character due to its own CLIP influence.
@zalupamirok696 That's great advise, thank you.
But yeah, I tried lowering the weight first, and lost identity then with the first version. That's why I developed an identity preserving img2img workflow around the two Krea 2 Edit nodes and the Krea Identity LoRA. As you can see with Yara v2 here, this does awesome magic.
@SenDai_Sempai I just tried your second version, and I have bad news for you: her hair looks like it’s nailed down—the hairstyle is impossible to change, even at a strength of 0.45. The side-swept look remains no matter what, which is a problem. Meanwhile, making her look significantly younger works without any issues, even at a strength of 0.7. The very thing you created the new version for isn't working, and now there's an issue with the hairstyle.
@zalupamirok696 well, it was not the intention to prevent age change completely, because those age modifying LoRAs will probably always enable that; but the issue with version 1 was that the default was already borderline, because the training material contained images that displayed her short stature of around 160cm. So the new dataset doesn't refer to size in any way.
The fixed haircut, though, is intentional with this model (in contrast to Blanka, for example). But probably I need to change the explanation in the text for that aspect. I forgot to mention that change.
Did you try mixing v1 and v2 together? How does this change the behaviour?
@SenDai_Sempai In short, it’s less flexible and relies more on its own "clip" (slightly more than the first one); it has a stronger influence on the seed, which is a downside. Changing the age is a breeze—you just specify the exact age; there’s no getting around that, buddy. As for height, 90% of women on Earth are on the shorter side—what’s wrong with that?