<h1 id="description">Description</h1>A Lora of my OC Mellanie. Her depiction in a prompt was quite elaborate, which led to inconsistent results depending on a combination of models/loras/embeddings and the context of the prompt. Made a Lora to bake those in, and I could ditch 90% of the Loras I was using to control the way she looked (10 -&gt; 2).<h2 id="character-sheet">Character sheet</h2><ul><li>Name: Mellanie</li><li>Age: 28</li><li>Temperament: Phlegmatic</li><li>Profession: Botanist</li><li>Occupation: Florist</li><li>Facial features: Thick lips and droopy eyes (tareme)</li><li>Body proportion: Tall, dark skinned, plump women with really generous female forms</li><li>Hobbies: Taking care of plants, going to the park and eating fast food.</li></ul><h1 id="versions">Versions</h1><h2 id="alpha">Alpha</h2>This is my first character Lora, so it's bound to have issues.If you find specific biases that is not described here, please let me know for future iterations to be rectified.<h3 id="training-dataset">Training Dataset</h3>65 images hand-picked generated on-site (t2i -&gt; i2i -&gt; facefix -&gt; highres -&gt; facefix -&gt; highres -&gt; upscale -&gt; facefix), manually color corrected (because generating dark skin is really hard on generic models, there is a bias against on average) and tagged.Most of the images are somewhat ok in term of representation, it's a best effort. Due to the initial prompt being really elaborate, results were not super consistent, specially for the face depending on angle. This might bleed into the Lora, but for what I've been testing, the results are pretty consistent, so I'm happy about that.All images had a simple white background, so there should be no influence on backgrounds in general.(40k buzz down the drain for 7 days of manual work in total... sigh)<h3 id="prompt-recommendation-for-canonical-representation">Prompt recommendation for canonical representation</h3>The Lora has a pretty good idea already, but some features were not well retained, so if you wish to approach more of the canonical representation of the character and address some of the short-comings of the Lora, considering adding those to your prompts:Weight: 0.8~1 (I use 1)Positive: <code>mature women, dark skin, tall, aqua eyes, tareme, thick lips, cute face, very long hair, plump, huge breasts, sagging breasts</code>info: without the breasts modifiers, the Lora outputs usually a good <code>huge breasts</code> (usually a bit bigger than the size of the head) equivalent with a pretty nice hourglass shaped body, for those that prefer a more 'modest' bossom and proportions. All showcased images are without breasts modifier, so people can get a better idea.Negative: <code>simple background, covered nipples, white hair, blonde hair</code><h3 id="known-issues">Known issues</h3><ul><li>Lora seems over-fitted, so it retained a lot of the style, which is a mix of a few other style Lora, the most strongly pronounced is Moriminee and some realism with WAI as a base.</li><li>Strong biais for covered nipples, this is due to the training dataset. It tends to be present when any character has large breasts it seems. Consider adding <code>covered nipples</code> to your negatives.</li><li>Sometimes there is a change in hair color, turning to white or blonde, not sure why, maybe it's a bleed from the initial dataset, which has only simple white backgrounds, and it trained white on the hair color. I suggest adding <code>white hair, blonde hair</code> to your negative</li><li>Sometimes background might be too simple if not specified otherwise. Probably coming from the dataset, which all hide <code>simple background, white background</code>. Consider adding those to your negative, if you notice you have plain backgrounds.</li></ul>

Files with SHA256: 8873c77f7b96e369bad9313d01bb8ad7192592a7fbcfc171a86dd9445efc4c03