Gia — a gorgeous young Italian woman with long dark brown wavy hair, warm olive Mediterranean skin, large dark brown almond eyes, strong arched eyebrows, and a softly angular oval face with high cheekbones. Slim athletic build with natural curves.
This LoRA captures Gia's face and identity for consistent renders across poses, outfits, and scenes. v3.0 is a full retrain of the face LoRA on a new base model with a corrected training set — sharper skin texture, natural skin detail, and a stronger likeness in low light.
Trigger word: gia_face
Recommended strength: 1.0.
- Her old V30 value does not transfer to this base. 1.0 is the value measured for the v40 lineage and confirmed on this file — the strongest likeness and the cleanest skin. Lower strengths render a weak, generic face.
- Pushing higher gives a stronger likeness in difficult scenes, at the cost of increasingly over-smoothed skin.
A matched body LoRA (gia_body) is included for figure consistency — body trigger: gia_body, strength 0.5.
Works best with: Krea 2 base models (v40 lineage). Photorealistic, natural Mediterranean daylight, warm practical lighting.
Example prompt:
gia_face, a gorgeous young Italian woman with long dark brown wavy hair, warm olive Mediterranean skin, large dark brown almond eyes, slim athletic build, standing facing camera in warm natural daylight, photorealistic, natural skin
Negative prompt tips: protect the hair and eye colour by negating blonde hair, light hair, red hair, blue eyes and hazel eyes. Negate pale/fair/porcelain skin to keep her olive tone. Keep plastic skin / airbrushed skin / waxy skin in the negatives.
This is a fictional character. All images are AI-generated.
⭐ Enjoying Gia? A rating + 👍 on the previews makes a real difference to an indie creator. Grazie, and thanks for the support!
Description
> ✅ WHY v3 EXISTS — this is the point of the version. v2 rendered heavy dark pigment marks on
> Gia's face. The cause was found in her TRAINING DATA, not the render settings: v2's 40 asset
> captions all carried the phrase "muted skin, matte complexion, natural skin texture", which
> taught the LoRA that her skin is muted and matte — a weight-level instruction to suppress
> texture, and with zero freckle/pigment language to anchor it, that suppression surfaced as marks
> over the new base. Verified by counting: v2 = 40/40 captions contaminated; v3 = 0/40.
> v3 also renders its assets on the skin07 bake instead of stock crown, so the training
> distribution actually contains the detail the LoRA is asked to produce.
>
> The old workaround is gone. v2 required a permanent skin-mark negative block to look
> acceptable. v3 does not — the fix is in the weights now, so **no skin-mark negatives are
> required and the README no longer asks you to add them.**
> ✅ SHE HAS A BODY LoRA NOW. Gia was FACE-ONLY before this version.
> gia_body_v1_cyberv40 is her first: 46 images, 16 epochs, dim64/alpha64, trained on v40 to
> match her face LoRA. A genuine capability addition, not just a version bump.