This is a LoRa I decided to train on gpt-img-2. There's two goals: improve facial texture and increase body proportions. Ideogram by default can't move away from a slim body. This pushes it to be nearly as good as the 200 elo better gpt-img. It happened to be mainly Korean, I generated most of the images before Ideogram was even released.
Trigger
Use:Korean
Place it once near the start of the woman’s description:
Korean, woman with...
Because Korean is also a normal word understood by the base model, it may affect facial appearance and styling as well as the trained body shape.
Tags That Work Well
large/full projecting chest— 288 images, 93%narrow shoulders— 236, 76%narrow or sharply indented waist— 214, 69%compact short-waisted torso— 209, 68%pronounced underbust taper— 197, 64%broad or rounded hips— 163, 53%
Appearance and clothing
dark or black hair— 294, 95%long dark hair— about 177, 57%white clothing— 287, 93%fitted or taut clothing— 275, 89%fitted white shirt/blouse/top— 262, 85%high-waisted— 234, 76%mini skirt— 226, 73%button-down blouse— 202, 65%gray clothing— 201, 65%pleated skirt— 196, 63%gray pleated skirt— 141, 46%black sheer tights— 128, 41%black skirt— 86, 28%white crop top— about 62, 20%
Settings
pool,flooded interior, or visible water — 106, 34%night— 72, 23%apartment, bedroom, or simple indoor room — 71, 23%street, sidewalk, or crosswalk — 60, 19%flooded corridoror flooded room — 58, 19%indoor pool— 48, 16%wet clothing— 42, 14%alley, passage, or corridor — 35, 11%wet pavementand urban reflections — 26, 8%instant-photo border— 21, 7%
Poses and camera
three-quarter stance— 121, 39%high-angleoroverhead view— 100, 32%front-facingornear-front stance— 82, 27%selfie perspective— 27, 9%side profile— 27, 9%low-angle view— 13, 4%
Default dataset look
Korean, woman with long dark hair, a compact short-waisted torso, a large projecting chest, sharply indented waist, and broad rounded hips, wearing a fitted white short-sleeve blouse tucked into a gray high-waisted pleated mini skirt and black sheer tights, three-quarter stance.
Water theme
Korean, woman wearing a fitted wet white blouse and gray pleated mini skirt, standing waist-deep in an indoor pool or flooded corridor, cool fluorescent lighting, dark reflective water, three-quarter view.
Night street theme
Korean, woman with long dark hair wearing a fitted white blouse and high-waisted dark mini skirt, standing beside a crosswalk on a wet city street at night, reflective pavement, soft storefront lighting, smartphone photograph.
Bright indoor theme
Korean, woman in a fitted white crop top and high-waisted gray or beige mini skirt, standing in a bright apartment near a large window, soft daylight, high-angle smartphone view.
Uniform-like theme
Korean, woman with long black hair wearing a fitted white short-sleeve button blouse, dark neck ribbon, gray pleated mini skirt, and black sheer tights, near-front three-quarter pose.
Compact Tag String
Korean, long dark hair, fitted white blouse, gray high-waisted pleated mini skirt, black sheer tights, compact short-waisted torso, large projecting chest, narrow waist, broad rounded hips, three-quarter stance, high-angle smartphone photograph
For the strongest in-distribution result, combine the morphology with white fitted clothing, a gray pleated mini skirt, and either a pool/flooded interior or a wet nighttime street.
Recommended strength
Start around:
0.6-0.8
The LoRA can influence body proportions even without the trigger, so lower the strength if unrelated woman prompts change too much.
Basic prompt
Korean, woman with a compact short-waisted torso, a full projecting upper torso relative to a narrow ribcage, a sharply defined waist, and rounded hips. Wearing a fitted blouse tucked into a high-waisted skirt, standing in a three-quarter pose.
Stronger morphology
Use more direct geometry:
Korean, woman with a compact short-waisted torso, massive breasts, deep cleavage and projection, strong upper-torso-to-waist contrast, a sharply indented natural waist, and broad rounded hips.
Important note
The LoRA does not act as a perfect on/off trigger.
Tests showed:
loading the LoRA can change body proportions even without
Koreanadding
Koreanmay also shift the face toward a Korean appearanceexplicit body descriptions still matter
For the strongest control, describe the body shape you want instead of relying only on the trigger.
Example structured prompt
{"aspect_ratio":"5:8","high_level_description":"A natural low-light smartphone photograph of a woman standing in a narrow rain-wet city alley at night, framed from above her hair to the upper thighs.","compositional_deconstruction":{"background":"A narrow brick service alley with closed shutters, exposed cables, small practical lamps, and wet pavement reflecting subdued cyan and amber light.","elements":[{"type":"obj","bbox":[20,70,1000,910],"desc":"Korean, woman with a compact short-waisted torso, very full projecting breasts relative to a narrow ribcage, a sharply indented natural waist, and rounded hips. Her breasts are extremely large. She wears a fitted white button-down blouse pulled taut across the upper torso and tucked into a dark red high-waisted pencil skirt. Near-front three-quarter stance, upright posture, both arms behind her back, long dark hair, and a calm direct gaze."}]}}
How It Was Trained
Base model: Ideogram 4 FP8
Training tool: ostris AI Toolkit
LoRA rank/alpha: 32/32
Training steps: 2500 but this one is 1750
Optimizer: AdamW 8-bit
Learning rate:
1e-4Resolutions: 512, 768, and 1024
Text encoder: not trained
Trigger word:
KoreanDataset: 309 retained image-caption json. Captioned manually by GPT-5.5 High with some guidance.
Captions used structured descriptions of body geometry, clothing tension, pose, camera perspective, and scene content. Mostly softened, not actually saying "huge breasts".
The dataset contains entirely synthetic Korean women from gpt-img-2, so the trigger may influence both morphology and facial appearance.
Description
Initial release. First training run. Great results but the dataset wasn't generated intending to train a model.



