V2 - Massive improvementTrigger Word: <code>Lelu</code> Use in combination with feminine pronouns (she, her) to reinforce the correct gender identity and expression.Model Type: WAN 2.1 14B LoRA Recommended Base: WAN 2.2Recommended Strength:<ul><li>High Noise: 0.6</li><li>Low Noise: 0.9</li></ul><hr /><h3 id="overview">Overview</h3>This LoRA introduces Lelu, a humanoid non-human female character originating from a sci-fi setting. The model reproduces her likeness, body proportions, and characteristic features with strong fidelity across a range of lighting and compositional contexts.She is designed to perform consistently in both realistic portrait and science-fiction environments, whether rendered in cinematic realism or stylized composition.<hr /><h3 id="prompt-guidance">Prompt Guidance</h3>Core Trigger:<code>Lelu</code>Optional Contextual Modifiers:<code>futuristic</code>, <code>sci-fi</code>, <code>starship</code>, <code>techsuit</code>, <code>workshop</code>, <code>spaceport</code>, or other science-fiction terms to steer the environment and lighting.Outfit Prompts: - (Better results and more options with V2)<ul><li><code>Lelu wearing a futuristic techsuit outfit, form-fitting long sleeve maroon and brown leather crop top that accentuates her curves with deep cut cleavage, paired with purple shorts and a utility belt adorned with various gadgets., brown leather gloves, a futuristic electronic device bracer on her right forearm</code> - Change the description up for variation but this will get you her standard look.</li><li><code>Lelu nude with heart-shaped sequins and rhinestones covering her body</code> – you can vary the color or shape of the sequins for creative flexibility.</li><li>wearing a metallic gold dragon-like decorative breastplate that adorns her chest, chainmail skirt</li><li>partially sheer embellished dress adorned with sparkling gemstones that catch the light</li><li>opalescent holographic lingerie</li><li>black leather harness criss-crossed style on her chest with open cutouts to expose her naked breasts</li></ul><hr /><h3 id="character-details">Character Details</h3>Lelu is a humanoid female with distinct non-human traits:<ul><li>Ears: long, drooping lop-ears.</li><li>Facial features: small feline-like nose and prominent upper canines.</li><li>Markings: animal-like stripe pattern visible along the upper ridge of her ears, across shoulders, and down portions of her arms, hips, and legs.</li><li>Eyes: green.</li><li>Hair: brunette, often styled with decorative ornaments, occasional braids, and a set of horizontal cornrows above the left ear (sometimes mirrored on the right).</li></ul><hr /><h3 id="notes">Notes</h3><ul><li>The LoRA includes full-body anatomy and nipple detail for accurate figure rendering.</li><li>Genital anatomy may require an additional NSFW LoRA for improved accuracy or realism.</li><li>Works with both High and Low noise conditions; adjust strength accordingly.</li></ul><hr /><h3 id="how-it-was-trained">How it was trained</h3><ul><li>113 dataset images were created for this LoRA.</li><li>I started with a handful of pieces of artwork of the character that either myself or friends have painted over the years and I used those as reference images for Seedream 4 4k to "bring her to life in a real photo". Initially I also used portrait photo images of random people as a style transfer reference image so that Seedream would more quickly converge on a photographic version of Lelu.</li><li>I created HUNDREDS of images and constantly had to manually edit and paint in specific features of hers that models don't tend to understand (like her ears and animal stripes).</li><li>Once I had enough facial likeness references, I was able to continue creating more images and even adding female anatomy images in as references to create those dataset images as well (see dataset images screenshot uploaded to the gallery below).</li><li>EVERY image in the dataset was edited in both Photoshop to remove blemishes or increase accuracy, and then in Lightroom to fix lip color, adjust skin detail, and smooth her ears (the AI kept generating them ultra textured so I had to make them match the rest of her body manually).</li><li>I captioned the images with BooruDatasetTagManager running<span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);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"> Qwen2.5-VL-7B-Instruct with a custom QUERY template for structured, vivid captions. For V2 - I manually edited each caption to add more detail and "prompt adherence"</li><li><span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);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">Trained in AI Toolkit with default WAN 2.1 14B LoRA settings (except you must change FPS and number of frames to "1") - 3,000 steps</li></ul>

Lelu - An original sci-fi character

V2 is a huge improvement. I went back to the same dataset and manually edited all my captions to be more accurate and more detailed and then retrained the model. This is something I normally would have done the first time but forgot, so I decided to be even more thorough this time and as a result, we get more interesting things we can prompt from my dataset as well as very good detail for her breasts and nipples (which were heavily featured in the dataset). Also, her genitals perform much better when accompanied by an NSFW model. The dataset features her genitals but in a more limited capacity so they still need some assistance.

Files with SHA256: a4884217b871433a879335332fc5b0b9c6145e3f667b2052889121896d9fcaaa