Krea 2 Piyali

If she works for you, a thumbs-up review helps other people find her.

Piyali is thirty-two, from North Kolkata: Coffee House mornings, Puja nights, a rooftop where she reads the paper and hangs the washing. Steady and amused, with a soft closed-mouth smile, a big laugh with the eyes creased, and a teasing look over the shoulder. A red-bordered white cotton saree on Puja days, printed cotton at home, silk at weddings, a kurti in town, a t-shirt and pyjama pants on the sofa. Fictional: she does not exist and no real person's images were used. Fourth woman from the same studio as Revathi, Thara and Ambili.

Run her

• Trigger: zxpiyali + her age: zxpiyali woman, 32 years old, then the scene.

• Krea 2 Turbo: strength 1.0, 8 steps, CFG 1, Euler / Simple, 1024×1344 (or 1152×1440).

• Sick Ollie and CyberRealistic Krea 2: strength 1.0, 9 steps, CFG 1; name her outfit in every prompt on these builds.

• Describe the scene, not her. Bindi, sindoor and the shankha-pola bangles are not her default: name them when you want them.

• Say "photograph", never "photorealistic". Keep readable print out of the scene.

Three prompts to start from

zxpiyali woman, 32 years old, waist-up, soft closed-mouth smile, hair in a low bun with white flowers, red-bordered white cotton saree, red bindi, old Kolkata balcony, soft daylight, photograph

zxpiyali woman, 32 years old, full body, walking down a narrow old Kolkata lane with a cloth bag, blue printed cotton saree, morning sun, 35mm, photograph

zxpiyali woman, 32 years old, phone photo, waist-up, laughing hard with a hand over her mouth, sitting on ghat steps, checked cotton shirt, overcast daylight, slight grain, photograph

Known limitations

• Without the trigger she still leaks in: a plain "32-year-old Bengali woman" prompt renders her as strongly as the trigger does (0.86 on the evaluation's two prompts without the trigger, and 0.86 for the same two scenes with it). For a different woman, unload the LoRA.

• True side profiles are her weakest framing (0.69 on the evaluation).

• Full-body wide shots hold her face less closely than close framings (0.76 against 0.88 for close-ups on the evaluation).

Measured, not claimed

Trained on 68 pictures of one fictional woman, rank 32 / alpha 32 on Krea 2 Raw (checkpoint step 2250 of a 2,500-step run), served on Turbo. Identity on an 18-prompt evaluation (16 prompts with the trigger, 2 without) against five reference pictures kept out of the training data: mean over the 16 trigger prompts 0.817, all 16 above 0.60 (close-up 0.88, waist-up 0.85, full body 0.76, damp hair 0.88, profile 0.69). A constant-frame test on other Krea 2 builds follows in an update to this card.

Derivative of Krea 2 under the Krea 2 Community License; not an official Krea product.