<h1 id="pal-076-wixen-kbcjay9lj">Pal #076, Wixen</h1><p>Aloof, afloof, and aflame. Sample pictures are all <strong>cherry-picked raw gens</strong> to demonstrate the model's capabilities in pure text2image. No inpainting or upscaling was done.</p><p></p><h1 id="prompting-rr7q86thv">Prompting</h1><p><strong>Weight:</strong> 1.0 works great. It can go as low as 0.7, but I haven't found many reasons to lower it.</p><p><strong>Trigger words:</strong> <code>wixen, semi-anthro</code> and <code>wixen, anthro</code></p><p><strong>For that sweet, sweet canon look:</strong> <code>wixen, semi-anthro, ambiguous gender, mouthless, huge tail</code></p><p><strong>For da boobies:</strong> <code>wixen, anthro, female</code></p><p><strong>Dudes too:</strong> <code>wixen, anthro, male</code></p><p>Also does ferals and taurs!</p><p></p><h2 id="reminder-tags-srusuxwzm">Reminder Tags</h2><p>In harder art styles, or more complex situations, the model might forget certain stuff. I've compiled a list of the tags you can use to guide it back to proper wixenity.</p><ul><li><p><code>3 fingers</code> - Especially useful when you've specified a hand gesture. It tends to make enough fingers to perform the gesture naturally; this'll encourage it to figure out how to do it with wixen hands.</p></li><li><p><code>red hat</code> - When things get extra weird, it can just straight up forget that it's drawing wixen. Putting a red hat on usually reminds it what it's doing.</p></li><li><p><code>mouthless</code> - The in-game model doesn't have a mouth! Weird, but true. I think they're adorable that way, your mileage may vary. A lot of the training images put a mouth on them, so you'll especially see it if you prompt an expression.</p></li><li><p><code>red nose</code> - Occasionally, making them mouthless also makes them noseless.</p></li><li><p><code>huge tail</code> - Fan art severely undersells just how enormous that bundle of fluff is. Poof it up a bit.</p></li><li><p><code>orange chest tuft</code> - Not technically a known tag, but if the front floof is coming out too white, this'll help.</p></li><li><p><code>red clothing</code> - When using a different model, the first thing that tends to go wrong is that the outfit turns blue. No idea why.</p></li></ul><p></p><h2 id="other-useful-tags-nn9lnaq3c">Other Useful Tags</h2><ul><li><p><code>alternate costume</code> - Makes it more flexible on the clothes. I didn't actually have any training data with alternate costumes, but it works anyway.</p></li><li><p><code>ambiguous gender</code> - Further discourages giving it a breast-like swell on the chest of the robe, if your art style or model are pushing it that way.</p></li><li><p><code>featureless chest</code> - If you want to generate naked wixen ladies without going full anthro, this'll keep more of the canon body shape.</p></li></ul><p></p><h2 id="wixen-noct-3e47ed3te">Wixen Noct</h2><p>Not actually trained on any, but try <code>alternate color, purple dress, purple hat, pink inner ear fluff, violet chest tuft, purple fire</code>. The front fluff is the hardest part to get right. I'll try to gather a few samples if I do another training run.</p><p></p><h1 id="quality-indicators-gveof5v7h">Quality Indicators</h1><p>Obviously, these are just my subjective ratings. But after testing a whole bunch, these are my experiences.</p><p></p><h2 id="character-likeness-6s2du02yd">★★★★★ Character Likeness</h2><p>It makes wixen. It does it with very little in the way of additional hints. I'm really happy with how it came out. Even the marks under the eyes come out close or correct.</p><p></p><h2 id="flexibility-wvb80llc2">★★★★☆ Flexibility</h2><p>Has no problem with different poses and situations. Doesn't seem to have dampened the model's ability to show emotion, and can even show a lot of emotions through the eyes and body language with the mouthless look. Can alter the outfit with some convincing. If it's being belligerent about an unusual pose, like handstands, you can lower the weight to 0.9 to loosen it up a little.</p><p></p><h2 id="style-compatibility-c3wsamn1i">★★★★★ Style Compatibility</h2><p>Styles remain distinct even at weight 1.0. I haven't noticed any significant style changes between applying it at 0.7 and 1.0. Lowering the weight can allow the art style to influence the <em>shape</em> of the character more, for example when trying to do bkub style. It can do realism on realistic models.</p><p></p><h2 id="other-checkpoints-f2bkw5d8c">★★★★★ Other Checkpoints</h2><p>I've tested it against a couple different versions of NoobAI, Willy's Noob Realism, and even my weird merge that's got 25% of an Illustrious-based model in it, and it transitions fine. It actually <em>really</em> likes NoobAI vpred 1.0, or that model's just smarter, because I get far better results from that than the model I trained it on. Surprisingly, it kind of works on Illustrious v0.1 if you load it at like 1.3+ weight and tag a lot of its traits. But that would be silly.</p><p></p><h2 id="other-characters-0ubsz4a81">★★★★☆ Other Characters</h2><p>I'm deducting a star because it's a LoRA and there's no way a LoRA isn't going to interfere with other characters at least a little. But in my tests with the LoRA loaded without its trigger word, I've been able to generate a mouthless arctic fox with a red nose, blue eyes, and a red hat, and the hat doesn't even look witchy. Generates multi-character scenes with Regional Prompter / Forge Couple.</p><p></p><h2 id="sanity-01b476l7l">★★★☆☆ Sanity</h2><p>Depending on model, sometimes it generates a weird blob of nothing wixen-looking at all. Also, when it gets overloaded, it can forget it's generating a wixen and make something else instead. A few reminders (see the prompting section) can usually fix that. I suspect the total image corruption quirk is because I used crap regularization data, and I might try to fix that with another training run later.</p><p></p><h1 id="training-info-o622iv3nk">Training Info</h1><p><strong>Base model:</strong> NoobAI-XL EPS 1.0</p><p><strong>Concept data:</strong> 27 human-created images with various styles and aesthetics (8/27 porn, 4/27 with breasts, 17/27 with mouth)</p><p><strong>Regularization data:</strong> 270 AI-generated images from the base model, same captions as the concept data, without "wixen", with randomized aspect ratios and quality tags</p><p><strong>Balance:</strong> 1 regularization image per 1 concept image, same loss weight</p><p><strong>Total Steps:</strong> 5000</p><p><strong>Notes:</strong> It was pretty solid on the likeness by 2000 steps, but it kept getting better without corrupting art styles or leaking into non-wixen anthros, so I let it keep going. Earlier versions would randomly make the outfit blue, or green, or black. Weird stuff.</p><p></p><h1 id="things-nobody-asked-(yet)-11pov1dxl">Things Nobody Asked (Yet)</h1><h2 id="why-there-are-already-so-many-wixen-loras...-h9gnz6wv5">Why? There are already so many wixen LoRAs...</h2><p>Mostly because I just like making my own stuff! Honestly, I didn't even test the other ones, and they might be even better than mine. But I liked this one so much I wanted to share.</p><p></p><h2 id="no-mouthsanthrossex-in-the-sample-images-cqrpfdmym">No mouths/anthros/sex in the sample images?</h2><p>I wanted to focus on the ability to produce their canon look, because that's the toughest part (and cute), and the flexibility across harder variables like situation, art style, and outfit. But don't worry, it knows where to put all the other stuff too. You can absolutely make porn with this LoRA.</p><p></p><h2 id="but-what's-under-the-hat-41wp4ir5s">But what's under the hat?</h2><p>Beats me. You can prompt it off, if you try hard enough, but it really confuses the LoRA. I think it's part of their head. The tip of the hat is a vent for excess heat.</p>

Files with SHA256: a998b2bd3ffbbced39cef39897a5c23de1c76980ae6defafa6fccc8921fb57e4