<h1 id="heading-484">Wilykit v3 - Thundercats 2011</h1>v3's preview images make use of the <a target="_blank" rel="ugc" href="https://github.com/Bing-su/adetailer">adetailer extension</a> to generate and inpaint the face at a higher resolution.This is a Lora based off of (you guessed it), Wilykit from the Thundercats 2011 reboot. v3 Is a complete rework of the training process that should provide a more flexible, robust, and accurate recreation of the character.TL:DRUse<pre><code>Wilykit, tail </code></pre>for the trigger wordsFor her default outfit (remove as necessary depending on composition)<pre><code>wkoutift, crop top, single strap, armlet, midriff, single vambrace, utility belt, belt pouch, skirt, barefoot, detatched leggings</code></pre>I recommend a weight of 1.5 to start off with (due to the training method used, a higher weight than usual is necessary). For a more human look, 0.8 will work fine, but you'll lose out on the facial markings, and body shading, but gain extra flexibility. You can push things as high as 2.5-2.8 without artifacting. Higher weights = more defined markingsThe lora should be flexible enough to work with most any model, and style lora. It even seems to work okay on realistic models!TL:DR Over, here's the more in-depth information!Here's what's changed in the process from v2 to v3.<ul><li>Retrained from scratch with a vastly expanded dataset. v2 had about 40 total images consisting only of fan art, v3 has over 170 split between fan art and official art. This total doesn't include head crops, but if it did, the dataset would be around 250+ total. The training method was the same as done for my Mereoleona Lora (see its description for details on the training process, or read the bit below).</li><li>Tagging was done manually for all images, as opposed to using auto-generated tags like the previous version.</li><li>Wilykit's signature outfit was specifically trained for and tagged. While also making sure that the flexibility is there to both put her in different outfits, but also make sure that the colors present in her default outfit does not bleed into prompted outfits.</li></ul>There's not really any noticable quirks present from the previous version (default outfit color bleed when not prompted for, style bleed from training images, tail not showing up, face/body markings not showing up)The only real thing that I've noticed is that the tail can come detached from the body from time to time, sometimes the tail isn't accurate, and sometimes multiple tails can show. From my experience with image generation though, this is par for the course and that tails hardly generate correctly even without the lora.Color bleed from the default outfit into other outfits should (for the most part) be less of an issue. There's still some instances where the colors will show up if you're not prompting a different color, but colors should be much easier to change this time around, so you shouldn't have to fight as much to get her to wear something other than dark blue/purple when you're prompting for a different color.Style bleed should no longer be an issue, and from my observations it should be pretty neutral in terms of styling. Some tags can shift things slightly, but it should stay much more neutral than the previous lora.As said above, due to the training method used, a higher weight than usual is necessary to bring out the finer details present in the character. If things like the facial markings, and body shading (white stomach, tan body, with the forearms and bottom of the legs white), aren't too important to you, then you're fine using a normal weight. The model is trained in such a way that higher weights shouldn't impact the style at all, since one of the primary objectives of the training method used was to minimize issues like style bleed (which was an issue with the previous model).Because of this, the higher weights can have a positive impact on your images, in that higher weights actually make the body markings stronger and more defined, with more contrast between them. Despite needing higher weights to run the lora in general, I personally see this as a win, since it essentially makes the body markings a switch that you can turn on and off without fundamentally altering the character.If you're interested in the training method (since I've been going on and on about it), it's the exact same as the one used for my Mereoleona Lora. I'll be writing a more in-depth guide in the near future, but here's the quick rundown.<ul><li>Make a lora using official art only (<a target="_blank" rel="ugc" href="http://fancaps.net">fancaps.net</a> is your best friend).</li><li>Make another lora using official art only, but train it only images you've cropped to the head.</li><li>Merge these two at a 0.7:0.3 ratio (Body at 0.7, Head at 0.3)</li><li>Do the same thing again, but using Fan art only.</li><li>Merge the final Fan art lora with the final official art lora at, again, a 0.7:0.3 ratio (fanart at 0.7, official at 0.3)</li></ul>The resulting final lora should be better than the sum of its parts, and provide a flexible character model, while also retaining details intrinsic to the character (due to the inclusion of the always on-model official artwork).The only tradeoff is that since it's a merge of 4 loras, the overall strength of it gets diluted, so your prompting weight has to get pumped up to compensate (hence why we use 2.0 here).I've made experiments of merging using higher, slightly overtrained epochs (overtrained if you were to use them as-is), with the hopes that it all averages out when merged, but so far the results seem to be negligible.

Wilykit v3 - (Thundercats 2011)

Retrained from scratch with a vastly expanded dataset. v2 had about 40 total images consisting only of fan art, v3 has over 170 split between fan art and official art.Tagging was done manually for all images, as opposed to using auto-generated tags like the previous version.Wilykit's signature outfit was specifically trained for and tagged. While also making sure that the flexibility is there to both put her in different outfits, but also make sure that the colors present in her default outfit does not bleed into prompted outfits.

Files with SHA256: 5e4184633b7441b23d4d12ae5ed2bbd7acb7d1c1d29a85603f596fd908c45676