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.<h2 id="trigger">Trigger</h2>Use:<code>Korean</code>Place it once near the start of the woman’s description:<code>Korean, woman with...</code>Because <code>Korean</code> 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<ul><li><code>large/full projecting chest</code> — 288 images, 93%</li><li><code>narrow shoulders</code> — 236, 76%</li><li><code>narrow or sharply indented waist</code> — 214, 69%</li><li><code>compact short-waisted torso</code> — 209, 68%</li><li><code>pronounced underbust taper</code> — 197, 64%</li><li><code>broad or rounded hips</code> — 163, 53%</li></ul>Appearance and clothing<ul><li><code>dark or black hair</code> — 294, 95%</li><li><code>long dark hair</code> — about 177, 57%</li><li><code>white clothing</code> — 287, 93%</li><li><code>fitted or taut clothing</code> — 275, 89%</li><li><code>fitted white shirt/blouse/top</code> — 262, 85%</li><li><code>high-waisted</code> — 234, 76%</li><li><code>mini skirt</code> — 226, 73%</li><li><code>button-down blouse</code> — 202, 65%</li><li><code>gray clothing</code> — 201, 65%</li><li><code>pleated skirt</code> — 196, 63%</li><li><code>gray pleated skirt</code> — 141, 46%</li><li><code>black sheer tights</code> — 128, 41%</li><li><code>black skirt</code> — 86, 28%</li><li><code>white crop top</code> — about 62, 20%</li></ul>Settings<ul><li><code>pool</code>, <code>flooded interior</code>, or visible water — 106, 34%</li><li><code>night</code> — 72, 23%</li><li><code>apartment</code>, bedroom, or simple indoor room — 71, 23%</li><li><code>street</code>, sidewalk, or crosswalk — 60, 19%</li><li><code>flooded corridor</code> or flooded room — 58, 19%</li><li><code>indoor pool</code> — 48, 16%</li><li><code>wet clothing</code> — 42, 14%</li><li><code>alley</code>, passage, or corridor — 35, 11%</li><li><code>wet pavement</code> and urban reflections — 26, 8%</li><li><code>instant-photo border</code> — 21, 7%</li></ul>Poses and camera<ul><li><code>three-quarter stance</code> — 121, 39%</li><li><code>high-angle</code> or <code>overhead view</code> — 100, 32%</li><li><code>front-facing</code> or <code>near-front stance</code> — 82, 27%</li><li><code>selfie perspective</code> — 27, 9%</li><li><code>side profile</code> — 27, 9%</li><li><code>low-angle view</code> — 13, 4%</li></ul><h3 id="default-dataset-look">Default dataset look</h3><code>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.</code><h3 id="water-theme">Water theme</h3><code>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.</code><h3 id="night-street-theme">Night street theme</h3><code>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.</code><h3 id="bright-indoor-theme">Bright indoor theme</h3><code>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.</code><h3 id="uniform-like-theme">Uniform-like theme</h3><code>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.</code><h2 id="compact-tag-string">Compact Tag String</h2><code>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</code>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.<h2 id="recommended-strength">Recommended strength</h2>Start around:<ul><li><code>0.6-0.8</code></li></ul>The LoRA can influence body proportions even without the trigger, so lower the strength if unrelated woman prompts change too much.<h2 id="basic-promptkorean-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.">Basic prompt <code>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.</code></h2><h2 id="stronger-morphology">Stronger morphology</h2>Use more direct geometry:<code>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.</code><h2 id="important-note">Important note</h2>The LoRA does not act as a perfect on/off trigger.Tests showed:<ul><li>loading the LoRA can change body proportions even without <code>Korean</code></li><li>adding <code>Korean</code> may also shift the face toward a Korean appearance</li><li>explicit body descriptions still matter</li></ul>For the strongest control, describe the body shape you want instead of relying only on the trigger.<h2 id="example-structured-prompt">Example structured prompt</h2><pre><code>{"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."}]}}
</code></pre><h1 id="how-it-was-trained">How It Was Trained</h1><ul><li>Base model: Ideogram 4 FP8</li><li>Training tool: ostris AI Toolkit</li><li>LoRA rank/alpha: 32/32</li><li>Training steps: 2500 but this one is 1750</li><li>Optimizer: AdamW 8-bit</li><li>Learning rate: <code>1e-4</code></li><li>Resolutions: 512, 768, and 1024</li><li>Text encoder: not trained</li><li>Trigger word: <code>Korean</code></li><li>Dataset: 309 retained image-caption json. Captioned manually by GPT-5.5 High with some guidance.</li><li>Captions used structured descriptions of body geometry, clothing tension, pose, camera perspective, and scene content. Mostly softened, not actually saying "huge breasts".</li></ul>The dataset contains entirely synthetic Korean women from gpt-img-2, so the trigger may influence both morphology and facial appearance.

Korean Body enhancer | Ideogram

Initial release. First training run. Great results but the dataset wasn't generated intending to train a model.

friendlyIdeogram4_v11.json

Korean_000001750.safetensors

optimizer.pt

v1.0

Welcome to my 💫🌄 Friendly Ideogram 4✨ Less mess, more magic and controlFeel like a real designer with Ideogram 4 and my friendly workflow!I offer my workflow with LLM Enhancer &amp; Prompt Builder modes. Compose each part of your generation!The latest version of ComfyUI is required.💻 System requirements:<ul><li>Minimum system requirements:</li></ul>RTX 3000-s, 8GB+ VRAM, 32GB+ RAM, 8-core processor, SSD, latest ComfyUI<span style="background-color:rgb(26, 27, 30);color:rgb(250, 176, 5);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">📌 Detailed tips and links to models in the workflow✨<span style="background-color:rgb(26, 27, 30);color:rgb(141, 141, 143);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px"> Workflow features:<ul><li>Extremely user-friendly interface</li><li>Maximum performance even on low-end systems with at least 8GB of VRAM</li><li>LLM Enhancer and Prompt Builder implementation</li><li>Support for up to 2 Lora</li><li>Detailed tooltips and links to models</li><li>Manual random seed for complete control over generations</li></ul>Model overviewIdeogram 4.0 is a powerful, open-weight text-to-image model designed for professional-grade visual creation. It features a 9.3 billion parameter single-stream Diffusion Transformer (DiT) architecture that was trained entirely from scratchKey Capabilities &amp; Features<ul><li>Production-Grade Typography: It offers crystal-clear, dense text rendering across multiple languages directly inside the composition, rather than as a fragile overlay</li><li>Native Transparency: The model can generate images with native transparent backgrounds, eliminating the need for post-processing background removal</li><li>High Resolution &amp; Flexibility: It supports native 2K resolution output and flexible aspect ratios up to 2048 pixels on each side</li><li>Advanced Controllability: The model is trained on structured JSON captions, providing unprecedented control over composition, style, lighting, and color palettes. It also allows for precise object and text placement using bounding box controls</li></ul>.Post-Generation Editing SuiteIdeogram 4.0 introduces a comprehensive suite of tools to revise and refine images after generation without starting from scratch:<ul><li>Layerize Text: Turns generated typography into editable layers so you can restyle, reposition, and refine the text afterward.</li><li>Prompt Edit: Allows users to describe changes in plain language to revise specific parts of the image.</li><li>Extend &amp; Reframe: Grows the composition beyond its original crop or shifts it for different aspect ratios while maintaining the original focal point.</li><li>Magic Fill: Adds new objects or replaces missing areas to seamlessly match the surrounding scene.</li><li>Other Tools: The platform also includes native background removal, upscaling for sharper delivery, and remixing to explore new creative directions.</li></ul>🤗🙏🏼 Thanks to <a target="_blank" rel="ugc" href="https://ideogram.ai/">Ideogram AI</a> and <a target="_blank" rel="ugc" href="https://github.com/kijai">Kijai</a>Original repo — <a target="_blank" rel="ugc" href="https://github.com/ideogram-oss/ideogram4/tree/main">GitHub</a>

Friendly Ideogram 4

<ul><li>NEW v1.1.1 from 22.06: added SageAttention function and links, feel 2x acceleration! </li><li>Added Image reference for LLM function - get JSON prompt from any image and use it in Prompt Builder</li><li>Added shift parameter for quality extra management</li><li>Fixed some bugs</li><li>Updated tips</li></ul>

Files with SHA256: 675e96556dfbfcf180b752d116b023c675c1972da8cb86c8777d34a49717e43b